Skip to main content
NCBI home page
Journal of Pathology Informatics logoLink to Journal of Pathology Informatics
. 2019 Sep 16;10:28.

Abstracts

PMCID: PMC6767787
J Pathol Inform. 2019 Sep 16;10:28.

A Novel Method for High Parameter Multiplex Phenotyping and Analysis of the Tumor Micro-Environment

T Regan Baird 1, Benjamin Freiberg 1

Background: Understanding the immune response to tumors has proven to be beneficial in the assessment of patient prognosis and selection for targeted immunotherapies. The tumor micro- environment (TME) is home to a large diversity of cell types that are identified by their biomarker signature. The phenotype of these cells are complex requiring the presence of several distinct biomarkers. More importantly, the presence or absence of certain cell phenotypes in the TME can be an indicator of the method of immunosuppression/immuno- activation within the TME. Therefore, the ability to accurately identify and classify cellular phenotypes in the TME is of growing importance to the overall understanding of cancer and disease progression. Typical immunostaining labels between 1-5 specific biomarkers on any tissue section; however, recent advances in fluorescence unmixing and imaging mass cytometry (IMC) have substantially increased the number of biomarkers that can be identified simultaneously. This allows for the discrimination of many more cell phenotypes that can be observed within the TME. The increase in phenotypic specificity and sensitivity has great potential to better decoding the TME and more accurately understand a patient's prognosis and enabling precision medicine based treatments.

Methods: Manually phenotyping cells can be an arduous and biased task. This project utilizes a novel software algorithm to cluster and reclassify individual cellular objects found in both fluorescence and IMC samples in a fully unsupervised manner. In addition, dimensional reduction using t-SNE plots was applied in order to observe phenotypic differences in healthy versus diseased tissue and in tumor versus stromal tissue compartments.

Results: The results of this study demonstrate that the unsupervised clustering and classification method employed had excellent accuracy and precision in defining cell phenotypes within both IMC and fluorescent images. The variable expression of phenotypes in healthy versus normal tissue and in tumor versus stroma is easily depicted in the analyzed images, tSNE plots, and phenotypic contribution charts.

Conclusion: The novel software algorithm utilized in this study demonstrates the potential for automated phenotyping of high parameter data in cancer research. In addition, the unbiased algorithm utilized has the potential to discover new cell phenotypes as it operates unsupervised and unbiased from without user input.

J Pathol Inform. 2019 Sep 16;10:28.

Large-Scale Implementation of Whole Slide Imaging for Primary Diagnostics in Anatomic Pathology

Jae-Hoon Chung 1, Anil Parwani 1

Background: Whole slide imaging (WSI) is a technology that allows a pathologist to visualize and analyze digitized slides. WSI can advance the field of pathology in various ways such as increased level of collaboration between pathologists, enhanced diagnostic accuracy via artificial intelligence, and easier access to stored images. In this paper, we outline the entire process of whole slide scanning and analysis by pathologists.

Methods: WSI technology was implemented in the Department of Pathology at the Ohio State University Wexner Medical Center (OSUWMC) for primary diagnostics and storage of digitized slides.

Results: Since May 2017, FDA-approved WSI technology has been used at the OSUWMC, and so far, 83,913 cases and 962,584 slides have been scanned. Primary diagnoses made via WSI were no less accurate than those made via traditional light microscopy.

Conclusions: WSI is an emerging technology with the potential to revolutionize the field of pathology. Our experience shows that high-throughput WSI can be successfully carried out at the level of major academic medical center such as OSUWMC.

J Pathol Inform. 2019 Sep 16;10:28.

Validation of Whole Slide Imaging for Primary Surgical Pathology Diagnosis of Prostate Biopsies

Sangeeta Desai 1, R Vidya 1, Pavitra Subramanian 1, Akash Sali 1, Santosh Menon 1

Background: Whole slide imaging (WSI) is an important component of digital pathology which includes high resolution digitization of glass slides and their storage digital images. Implementation of WSI for primary surgical pathology diagnosis is evolving, following various studies which have evaluated the feasibility of WSI technology for primary diagnosis.

Methods: This was a single centre, observational study which included evaluation by 3 pathologists and aimed at assessing concordance on speciality specific diagnosis and comparison of time taken for diagnosis on WSI and CLM. Seventy prostate core biopsy slides (reported between January 2016 and December 2016) were scanned using Pannoramic MIDI II scanner, 3DHISTECH, Budapest, Hungary, at 20X and 40X. Sixty slides were used for validation study following training with 10 slides. As per CAP guidelines, intra-observer agreement between WSI and CLM was calculated, after a wash off period of 1 month. Intra-observer concordance for diagnosis between CLM and WSI was analysed using Cohen's κ statistics and intra-class correlation coefficient (ICC); Observation time for diagnosis were compared by paired t-test.

Results: Interpretation on WSI using 20x and 40x was comparable with no major discordance. High level of intra-observer agreement was observed between CLM and WSI for all 3 observers, both for broad diagnosis (κ = 0.9) and specific diagnosis (κ = 0.9). The intra-observer agreement between CLM and WSI for the Grade group (κ = 0.7-0.8) was good for all three observers. The major discordance between CLM and WSI was 1.6-6.6%, which reflected the expertise of the observers. The mean time taken to arrive at diagnosis for the Observer 1 and 2 was less on WSI as compared to that on CLM; whereas it was more on WSI for the Observer 3.

Conclusion: WSI is comparable to CLM and can be safely incorporated for primary histological diagnosis of prostate core biopsies.

J Pathol Inform. 2019 Sep 16;10:28.

Using GoToAssist to Improve Slide Scanning Workflow in a Distributed Setup

Robin Dietz 1, Douglas Hartman 1, Matthew O'Leary 1, Anthony Piccoli 1, Jon Duboy 1, Ian Karwoski 1, Alexander Mensing 1, Liron Pantanowitz 1

Background: Whole slide imaging scanners at UPMC were deployed at multiple locations. Scan technicians are therefore required to work at 6 different hospital locations. After loading slides into a scanner there is a lengthy time lapse (minutes to hours depending on the number of slides loaded) to generate a tissue snapshot to finalize scanning. Waiting for snapshots to generate is an inefficient use of scan technician time. Therefore, a better method to remotely monitor scanning from a central location was desired.

Technology: GoToAssist (version 4.5) was employed to remotely access Aperio AT2 whole slide imaging scanner consoles (Leica).

Methods: Glass slides were loaded daily into 7 Aperio AT2 scanners at different hospitals by two slide scanning technicians. GoToAssist was used to remotely monitor and troubleshoot scanning for all devices.

Results: GoToAssist allows multiple consoles to be monitored from a single location which cuts down on travel time. Scan technicians each report saving around 1 hour daily when using GoToAssist to remotely oversee the scanning operation at UPMC. Time-savings are derived mainly from allowing technicians to be able to remotely fix errors that delay digitization (e.g. slides marked red in the “Rack View” of the console that prevents scans if not resolved). GoToAssist thereby prevents a backlog of unscanned slides. Despite rare network connection problems, GoToAssist has functioned very well after around 1 year of use.

Conclusion: Using GoToAssist at UPMC improved the efficiency of our scan technicians, allowing them to remotely monitor multiple slide scanners from a central location. Utilizing the web-based application GoToAssist to remotely monitor and support the distributed whole slide imaging scanning setup at UPMC has limited the number of technicians required to scan our large volume of slides.

J Pathol Inform. 2019 Sep 16;10:28.

Performance of DeePathology. ai Helicobacter pylori Detection Software

Robin Dietz 1, Liron Pantanowitz 1, Douglas Hartman 1

Background: Helicobacter pylori (HP) organisms can be difficult to spot on hematoxylin and eosin (H&E)- stained sections and may require immunohistochemistry (IHC) for confirmation. DeePathology.ai offers a software program designed to assist pathologists in detecting HP organisms. We tested the software's performance to detect HP in 52 positive and 48 negative IHC-proven gastric biopsies.

Technology: DeePathology.ai software was downloaded from https://deepathology.ai. 100 H&E- and HP IHC-stained gastric biopsies were scanned at 20x and 40x magnification on Leica Aperio AT2 scanners.

Design: The time taken to run the software on each slide and the maximum detection threshold to detect organisms was recorded. The sensitivity and specificity at each detection threshold per H&E-slide was calculated and used to generate a ROC curve and calculate the AUC for each magnification.

Results: The AUC for 20x scans was 0.68 and 0.76 for 40x scans [Figure 1a]. We found the detection threshold of 0.475 to work best, which is consistent with that recommended by deePathology.ai to call a slide positive for organisms. At a detection threshold of 0.475, we found 13% false negatives and 17% false positives. While the detection of HP organisms by the algorithm correlated well with organisms seen in the IHC stains, it was not always consistent across tissue sections. Runtime increased with file size [Figure 1b]. 9% of slides scanned at 40x magnification failed to run. Conclusions: The software was able to identify some organisms, but not well, at the task detecting of HP organisms. Although it identified areas with organisms that would be difficult to detect on H&E sections, it was not consistent across serial tissue sections on the same slide. 20x and 40x scans performed similarly. DeePathology.ai may be a useful adjunct to H&E to assist in finding HP organisms, but is best used with pathologist oversight. Further testing of this algorithm as a clinical decision support tool is underway.

Figure 1.

Figure 1

Open in a new tab

(a) Receiver operating characteristic curve of ×40 and ×20 scans. (b) File size (MB) versus time to run deepathology. AI (s) color coded by threshold

J Pathol Inform. 2019 Sep 16;10:28.

Aiding the Surgical Pathologist: Automated Triaging and Screening Cases via Low-Power Images and a Microscopic-Field Based Machine Learning Approach; Colon Polyps for a Start

Lucy Fu 1, Uwadia Edomwonyi 2, Yash Dharmamer 2, Anjana V Yeldandi 3, RManmeet B Singh 2, Tushar Patel 2, Alaa Alsadi 2

Background: Pathologists devote a significant amount of time evaluating numerous minimally complicated biopsy specimens. Just as hematopathologists have increasingly adopted automated cell counters, surgical pathologists with selective high volume cases (colon polyps in our study) could achieve higher sign out efficiency. We aimed to design an accessible, pathologist-independent workflow to triage and predict findings in colon polyp biopsies using widely available hardware and open source software.

Methods: A prototype set of static images from unique, low magnification (4x and 5x) microscopic fields were collected via three image acquisition modalities with varying complexity and resolution: iPhone 6s with iDu Optics microscope adapter; ViewsIQ Panoptiq telepathology system; and Hamamatsu NDP Nanozoomer scanner (images exported for screen capture on an Apple Cinema Display). Image pre-processing and data augmentation (by 2x, macOS Preview), expert image labeling (Diagnosis Protocol), and neural network training (Microsoft Customvision.ai and NVIDIA DIGITS) were utilized to classify and predict the following pathologies: tubular adenoma, normal mucosa, hyperplastic polyp, sessile serrated adenoma, lymphoid aggregate, and other. We then evaluated the diagnosis- specific predictive probabilities for each training platform and image acquisition modality; Table 1.

Table 1.

Comparing precision (%) of NVIDIA and microsoft DL platforms trained on iPhone, panoptic, and Hamamatsu images

NVIDIA/Panoptic* Microsoft/iPhone Microsoft/Panoptic Microsoft/Hamamatsu
Tubular adenoma 65 83.4 83.9 89.1
Normal mucosa 65 45.8 50 49.7
Hyperplastic polyp 65 16.2 25.5 49.1
Sessile serrated adenoma 65 37.5 57.9 57.7
Lymphoid aggregate 65 44 46.8 47.4
Other 65 61.1 60 **
Open in a new tab

*NVIDIA digits provided an aggregate accuracy of 65%. The specific accuracy would likely decrease for diagnoses with fewer training images and increase for diagnoses with more training images, **Too few samples with “other” diagnosis with the Hamamatsu scanner to product a prediction

Results: We retrieved 117 biopsy cases from which 256 parts were from colon. We obtained 1711 unique microscope field images. Data augmentation by 2x produced a total of 3422 training images. The diagnostic predictive performances are reported in Table 1. Predictions for a sample test polyp for both machine learning platforms is presented in Figure 2.

Figure 2.

Figure 2

Open in a new tab

Diagnostic prediction for sample test polyp

Conclusions: Our workflow, which utilizes publicly-available machine learning platforms, could be implemented in any pathology practice today, to relieve high volume queues of minimally complex colon biopsy specimens. Specimens that have a high prediction for benign diagnosis can be deprioritized to allow time for more complex cases. Furthermore, pending cases could be assigned preliminary sign out reports with pre-filled diagnostic predictions that can be quickly verified/amended by a pathologist. Finally, future enhancement in diagnostic prediction by increasing the set of training images and involving expert personnel to fine tune custom neural networks will elevate the accuracy and applicability of this approach to real world.

Figure 1.

Figure 1

Open in a new tab

Proposed workflow utilizing widely available hardware and open source software

J Pathol Inform. 2019 Sep 16;10:28.

Life in the Fast Lane: Utilization of Telepathology for Remote Intra - Departmental Consultation

Yipeng Geng 1, Thomas Chong 1, Tobi Quinto 1, Stephen D Hart 1, Nicole Dawson 1, William D Wallace 1

Background: Many pathology groups are multi-centric with each location having different domain expertise in different areas of pathology. Intra-departmental pathology consultations may require labor and time-consuming glass slide transportation workflows. Our institution has one main campus in Westwood and two satellite locations in Santa Monica and Northridge, California. We explored using telepathology technology to facilitate rapid and workflow- friendly intra-departmental consultations and compared our findings with conventional courier- based glass slide delivery.

Technology: We have deployed two 5-slide Capacity Leica Biosystems CS2 scanners at our Santa Monica and Northridge hospitals and one 400-slide capacity AT2 scanner at our Westwood campus. All sites utilize Leica Biosystems eSlide Manager digital image management system to share the images within our department through an integration with our laboratory information system, Epic Beaker.

Methods: We performed a time study at all institutions to compare performance of the digital pathology workflow versus conventional courier transportation. We recorded both staff labor time and case transition time between both workflows over a 4-week period and compared the results.

Results: The average time for the slides to be physically transferred and be available for review to and from Northridge (16 miles from Westwood) was 437 min for 22 cases; and the average time to and from Santa Monica (4 miles from Westwood) was 431 min for 21 cases. In contrast, the average time for 10 digital cases for labor/scanning and be available for review was 22 min, ranging from 7 to 45 min depending on the number of slides scanned. The decrease in scanning and labor time versus courier time is statistically significant (p<0.01). Across all sites, there is an insignificant increase in average staff labor time between conventional transfer of slides and digital scanning (6.8 min and 8.4 min, respectively; p0.28). However, total glass slide transit time at the Westwood campus was significantly greater due to greater distance from pathologists' offices and the courier pick-up point (labor and transportation 32.3 min, p<0.00001).

Conclusions: Our results indicate substantial and statistically significant improvement in intra- departmental turn-around time but no statistically significant difference in staff labor time.

J Pathol Inform. 2019 Sep 16;10:28.

Expanding the Use of Hybrid Scanners beyond Remote Intraoperative Consultation

Douglas J Hartman 1, Kimberly Young 2, Sureshchandra Patel 2, Liron Pantanowitz 1

Background: Using virtual microscopy to support remote intraoperative consultation is a well described use case for digital pathology. With the introduction of hybrid scanners, these devices can be utilized for more than simply live intraoperative consultation. Herein, we describe our experience using a hybrid scanner for both remote frozen sections and post signout review of routine diagnostic cases.

Technology: Leica/Aperio LV1 device (Leica Biosystems, Buffalo Grove, IL) was used for scanning. Whole slide images were imported into the Aperio eSlide Manager system (Leica Biosystems, Buffalo Grove, IL) where the pathologist reviews were performed.

Design: We compared the use of digital slides for post signout (quality assurance) review of surgical pathology cases versus the prior method of sending glass slides more than 100 miles from the original site. 224 cases were reviewed: 126 cases using glass slides and 98 using digital slides. The time for review was calculated from the signout date until the date of the post signout review.

Results: The average time to post signout review was 58 days (digital method) versus 145 days (glass slide method). It takes approximately 13 minutes to prepare each case according to the glass slide method. Scan time for slides was approximately 13 minutes for each case. Although the preparation time is similar, the sending site is able to keep their slides and eliminate transportion costs of a courier.

Conclusions: With the introduction of hybrid scanners, the devices that have been deployed for remote intraoperative consultation can now also be used for other purposes. We found that using digital slides for post signout review shortened the average review time per case by nearly 3 months. For the sending site, the digital method was more convenient and secures the integrity of the slides. Multiple use cases can help to add to the return on investment of a hybrid scanner.

J Pathol Inform. 2019 Sep 16;10:28.

Utilizing Digital Image Analysis for Tumor Bud Quantification Demonstrates No Location Preference of Highest Tumor Buds and May Be a Helpful Adjunct to Tumor Bud Scoring

Patrick Henn 1, Reetesh Pai 1, Douglas J Hartman 1

Background: Tumor bud (TB) identification and quantification is an important adverse prognostic factor in colorectal adenocarcinoma. It is recommended to assess TB at the “invasive front” but no specific guidelines for where along the “front” have been established. This study aims to 1) determine where the highest TBs occur, and 2) evaluate the utility of digital image analysis in quantifying tumor budding.

Design: Twenty-four consecutive untreated colorectal adenocarcinomas were identified to include eight cases of low, intermediate, and high reported TBs. The TNM stage was recorded. All tumor slides were digitized. Using ImageScope Annotations, consecutive boxes with an area of 0.7849 mm2 were placed along the tumor front and TBs were manually annotated within each box. For each slide, the box with the highest TB count was identified and the location of that box designated as submucosal, muscularis propria, or subserosal. The TB score was compared with the reported score.

Results: 125 tumor slides were digitized and 3,380 boxes were placed with 1,717 containing TBs. 4/24 cases had designated the slide TB assessment was performed on. Nine of 24 cases demonstrated discrepancy between the highest field and the reported TB assessment. 3/8 cases with low TB score by report disagreed with digital assessment: one T3N0 case scored high and two cases (T3N0 and T4aN1a) scored intermediate by digital evaluation. 5/8 cases with intermediate TB score by report disagreed with digital assessment: one T3N0 case scored low and four cases (T1N0, T3N0, T4aN0, and T1N0) scored high by digital evaluation. 1/8t cases (T2N0) with high TB score by report disagreed with digital assessment (intermediate). For stage T3 and T4, the highest TB score was identified in the submucosa (6 cases), subserosa (8 cases), and muscularis propria (5 cases).

Conclusion: For stage T3 and T4, the highest TB count does not occur in a particular location along the invasive front. Digital assessment of TBs as compared to routine assessment resulted in a lower TB score in 2/24 cases (8.3%) and higher TB score in 7/24 cases (29.1%). Digital image analysis may be more accurate and a helpful adjunct in TB scoring.

J Pathol Inform. 2019 Sep 16;10:28.

Predictive Colorectal Cancer by Machine Learning

Ping Chong Ho 1, John Mok 1, Vincent Ng 1, Vicky Fung 1

Background: Colorectal cancer is the commonest cancer in Hong Kong for which faecal occult blood testing is a conventional screening test to justify colonoscopy referral. However, a predictive model using machine learning for early detection of colorectal cancer by analysing blood counts, age and sex that has been validated overseas. This is a proof of concept study on Hong Kong adult population that we were running machine learning algorithms to predict colorectal cancer by completed blood count (CBC) results, patients' age and sex from a general acute hospital.

Technology: Weka (Waikato Environment for Knowledge Analysis) developed by University of Waikato, New Zealand is used for supervised machine learning training and modelling. Labelled training and testing datasets saved as a comma separated variable file are curated from standardized Hospital authority laboratory data which are mapped to the reference terminologies – LOINC for laboratory tests and SNOMED CT for anatomical pathology terminology.

Methods: It is a cohort study with 264 positive colorectal cancer cases and 9444 negative cases. One year CBC data were extracted with 10 years anatomical pathology data. If patients do not have pathology investigation requested for 10 years before the CBC reporting date, the related CBC data could be classified as negative case. Whereas if there is any colorectal cancer pathology results reported within one year after the CBC results, the CBC data are classified as positive case. Descendant concepts of “Malignant neoplastic disease (disorder)” or “Malignant neoplasm, primary (morphologic abnormality)” from the SNOMED CT is used to identify malignant in anatomical pathology data.

Results: We found that running with those features under Random Forest with Cost Sensitive Classifier could produce the best accuracy of predictive modelling for the colorectal cancer – Area Under the Curve (AUC) is 0.814.

Conclusions: This local study shows that supervised machine learning to use patient's age, sex and CBC results can predict the likelihood of colorectal cancer.

J Pathol Inform. 2019 Sep 16;10:28.

Bag of N-gram Features as Supervised Classifiers of Cervical Pathology Specimens: A Retrospective Comparison of Cervical Cancer Screening Strategies

Jim Hsu 1, Paul Christensen 1, Yimin Ge 1, S Wesley Long 1

Background: Persistent high-risk human papillomavirus (hrHPV) causes precancerous lesions that progress to cervical cancer. Given high specimen volumes (>100,000 samples generated annually), accurately classifying cervical biopsy results remains challenging. Errors in classification adversely impact patients and providers.

Technology: Natural language processing (NLP) has been used to increase classification accuracy in myriad applications ranging from quality improvement in factories, anomaly detection in national security, and photo recognition on social networks. The introduction of distributed representations of word vectors indifferent to word ordering, as represented in the original Skip-gram model and further refined by biased sampling against frequent words (Word2Vec, Google) and capturing local word ordering (FastText, Facebook) has led to approaches that improve the scalability and usability of natural language processing in various applications, including internet search, image recognition, and content tagging.

Methods: We adopted a bag-of-words approach (Word2Vec) as well as a bag of N-grams approach (FastText) to generate an efficient baseline for natural language processing of cervical pathology specimens, comparing this approach to manually generated Regular Expressions, and evaluated the discriminative ability of a multivariate classifier for cervical biopsy free-text interpretation.

Results: We show that a bag of N-grams approach using an 80% training (N=48700) / 20% validation (N=11951) dataset has concordance comparable with previously generated regular expressions (precision@1 = 0.984; recall@1 = 0.984) for the correct classification of cervical biopsies (Benign, LSIL, HSIL, carcinoma, Blank/not interpretable). We additionally define measures for categorization quality, and explore cervical biopsy interpretations with low or borderline scores to define subsets of “problematic” interpretations for further study, illustrating some of these examples.

Conclusions: We established the utility of using natural language processing techniques to achieve accurate, consistent classification of cervical biopsy results which can be extended to categorical diagnosis for other specimens. We also identified potential difficulties in NLP-based classification, and foresee data from this approach could better inform pathologists when categorizing potentially problematic interpretations.

J Pathol Inform. 2019 Sep 16;10:28.

The UCLA Experience of Formalized Pathology-Radiology Correlative Reporting Using an Integrated Diagnostic Platform

Shohei Ikoma 1, Shawn Chen 2, Fereidoun Abtin 2, Scott Genshaft 2, David Lu 2, Steven Raman 2, Anthony Sisk 1, Robert D Suh 2, Corey W Arnold 2, W Dean Wallace 1

Background: The UCLA departments of pathology and radiology have developed a combined diagnostic workflow for pathology and radiology to synthesize and improve the final diagnostic report. This new service, called Integrated Diagnostic Report (IDR), has been deployed for the lung and prostate service lines since 2013. The IDR system is a web-based application built on Java- based Grails framework. It retrieves and integrates imaging and diagnostic data stored on Centricity picture archiving and communication system (PACS) (GE Healthcare, Chicago, IL) and Beaker laboratory information system on Epic EHR system (Epic Systems, Verona, WI). The IDR workflow begins when the pathology case has been finalized in the LIS. The IDR interface then prompts the radiologist to enter interpretive correlation comments. Once the r e p o rt is finalized, a hyperlink to the corresponding IDR is automatically e-mailed to the referring provider, immediately published in the patient chart, and linked to the original radiology and pathology reports.

Methods: To determine the overall effectiveness of the system, we audited the IDR database and examined the usage and clinical impact to determine the value of the system 5 years post- implementation. We evaluated the number of cases generated and the number of views per case for each service line. We also tracked changes over time to assess growth.

Results: Over the course of 5 years since its implementation, a total of 2,917 IDR cases have been completed, of which 1888 have been prostate and 1,029 have been lung. The average completion time was 30 and 127 seconds for pathologists and radiologists, respectively. There have been a total of 9,710 case views, for 3.3 views/case. The number of new cases per month have increased from 11 in Nov 2014 to a high of 112 in April 2018. The overall reception of the system by both clinicians and diagnosticians has been positive for its ease of use, ability to consolidate diagnostic data, and impact on multidisciplinary treatment planning. Currently, the IDR system serves as the presentation platform for the thoracic tumor board.

Conclusion: The IDR system facilitates highly contextualized communication among pathologists, radiologists, and referring clinicians. Improved coordination of the diagnostic process promotes efficiency and patient safety, providing significant value to the health system.

J Pathol Inform. 2019 Sep 16;10:28.

Fifteen Years of Image Analysis in a Clinical Setting: A Retrospective Analysis of the Evolution of Technologies and Procedures

Jennifer Jakubowski 1, Sharon Cavone 2, Mark D Zarella 1, Fernando U Garcia 2

Background: Adoption of whole-slide imaging in the clinical workflow faces many challenges. Among these, differences observed over time and across laboratories due to changes in whole- slide scanners, acquisition parameters, image platforms, and analysis algorithms have been cited as potential sources of variability. However, other factors, such as histologic preparation and staining protocols, are known to also contribute to variability. To distinguish between these factors, we examined whole-slide images, annotations, quantitative scores, and reporting obtained in a clinical capacity over the course of 15 years at a single institution. With the aid of detailed documentation highlighting significant changes to the imaging system (as required by CAP), we correlated a set of quantitative measures of image attributes to these events to compare their relative effects.

Methods: Using a large annotated de-identified data set collected over the course of 15 years, we examined the impact of changes in whole-slide imaging technologies and procedures within a single laboratory. This enabled an evaluation of the impact of technology on factors important for pathology diagnostics independent from the differences frequently observed across laboratories.

Results: Whole-slide image acquisition remained largely unchanged over the analyzed period with the exception of two major events: one in which magnification was increased from 20x to 40x and a second in which a new slide scanner was incorporated. Despite these changes, image quality and annotation precision exhibited remarkable consistency during these transitions. We examined the number of rescans and the number of out-of-focus regions identified in acquired H&E scans, and found a slight reduction following the incorporation of a new scanner into the workflow but not a significant change following the increase in magnification. In addition, we found that immunohistochemistry scoring remained consistent across transitional periods, suggesting that the scanner and acquisition parameters had a negligible effect on image analysis. Finally, we examined the color attributes of whole-slide images of H&E slides using an automated classification algorithm, and discovered significant changes unrelated to the imaging protocols and technologies, likely due to changes in histologic preparation and staining protocols.

Conclusions: These results underscore the relative impact of changes in laboratory protocols, equipment, and personnel, and demonstrate that imaging procedures that are sufficiently validated can be made to be robust to changes in technologies.

J Pathol Inform. 2019 Sep 16;10:28.

Digital Image Analysis Workflow for CD8+ Tumor Infiltrating Lymphocytes in NonSmall Cell Lung Cancer

Iny Jhun 1, Emilio Madrigal 1, Long P Le 1, Mari Mino-Kenudson 1

Background: CD8+ tumor infiltrating lymphocytes (TILs) and PD-L1 are reported to have important predictive relevance for PD-1/PD-L1 axis blockade in non-small cell lung cancer (NSCLC). While manual counting of PD-L1+ tumor cells is clinically applied for NSCLC, a scoring system for CD8+ TILs has not yet been established. Digital image analysis (DIA) has the potential to provide an efficient and standardized approach for estimating CD8+ TIL density. We aimed to test the feasibility of a DIA workflow for calculating CD8+ TIL scores in dual PD-L1/CD8 stained slides of NSCLC.

Methods: Sixteen NSCLC cases spanning multiple histologic subtypes (11 adeno-, 3 squamous cell, 1 large cell neuroendocrine, 1 pleomorphic carcinomas), tissue sites (e.g., brain, lung, lymph node), and specimen types (e.g., biopsy, resection) were retrieved and digitized at 40x magnification (0.2214 μm pixel width and height). Pathology reports for each case were exported. A DIA workflow was developed in QuPath (Belfast, UK) and comprised of stain estimation, cell detection, feature computation, tumor identification, and stain intensity classification. A random tree tumor classifier was trained using 8 cases. The estimated CD8+ TIL density was compared to scores in pathology reports by 7 pathologists, who semiquantitatively estimated the percentage of tumor cells associated with CD8+ TILs and assigned a score of 0, 1, or 2 with 5% and 25% as cutoff values.

Results: The DIA workflow applied to 16 dual PD-L1/CD8 stained slide images (mean area: 12,254,089 μm2) identified an average of 18,597 (61 - 75,919) cells in the tumor region including tumor cells and TILs. An average of 3.6% (0.0 - 36.8%) of tumor cells had CD8+ TILs on them. Estimated CD8+ TIL scores were in agreement with pathologist-reported scores in 7 out of 8 training cases and 4 out of 8 validation cases [Table 1]. For all cases with discrepancies in scores, estimated CD8+ TIL scores were lower by 1 compared to pathologist-reported scores.

Table 1.

Comparison of manual and image analysis CD8+ tumor infiltrating lymphocytes scores in 8 training cases and 8 nontraining cases

Tissue type Tumor type CD8+ TIL score
Pathologist DIA
Training
Brain Adenocarcinoma 2 2
Brain Adenocarcinoma 3 2*
Lymph node Large cell neuroendocrine 1 1
Lung Adenocarcinoma 1 1
Lung NSCLC, NOS favor squamous cell 1 1
Lung Adenocarcinoma 1 1
Lung Pleomorphic carcinoma 3 3
Lung NSCLC, NOS favor squamous cell 1 1
Validation
Brain Adenocarcinoma 2 1*
Brain Adenocarcinoma 2 1*
Lung Adenocarcinoma 1 1
Lung Adenocarcinoma 1 1
Lung Adenocarcinoma 1 1
Lung Squamous cell 1 1
Lung Adenocarcinoma 2 1*
Lung Adenocarcinoma 2 1*
Open in a new tab

Asterisks (*) indicate discrepancy in scores. TIL: Tumor infiltrating lymphocytes, NSCLC: Nonsmall cell lung cancer, DIA: Digital image analysis, NOS: Not otherwise specified

Conclusions: Our analysis demonstrates challenges in creating a generalized DIA workflow for various NSCLC tumor types. Given the heterogeneity in NSCLC morphology, tumor identification classifiers specific to histologic subtype will likely yield greater accuracy in tumor classification and CD8+ TIL scores.

J Pathol Inform. 2019 Sep 16;10:28.

Combining Deep Learning with Classical Domain-Based Detection for the Automated Identification of Mitoses

D Harding 1, N Verma 1, A Mohammadi 2, J Monaco 2, M C Lloyd 1, G Tozbikian 2, Z Li 2, A P arwani 2

Background: In patients with node-negative breast cancer, mitotic activity is highly prognostic of survivaltimes. Unfortunately, the manual identification of mitotic figures is both time-consuming and subject to significant inter-observer variability. Adoption of digital pathology facilities the development of automated image analysis algorithms to efficiently and reliably identify mitotic figures. Recent advances in machine learning techniques such as deep learning have significantly enhanced the performance of automated image analysis systems. By mining features directly from the image data, these deep learning systems do not require the direct embedding of domain knowledge (from experts, such as pathologists). We combined a traditional system that uses domain-specific features with a modern deep image segmentation network to see if such a hybrid system could outperform either.

Methods: We selected the deep learning network Retinanet (Facebook AI Research, Menlo Park, CA). Both the traditional system and hybrid systems were implemented and evaluated using MATLAB 2016b (Natick, MA). All images and annotation data were provided by the AMIDA13 MICCAI Grand Challenge; digitized was performed using an Aperio ScanScope XT at the University Medical Center Utrecht (Utrecht, Netherlands). Each system was individually trained on a training subset of the AMIDA13 data. The mitotic figures detected by each system were combined and reclassified using the confidence values reported by both systems. Performance data was gathered over the disjoint testing subset of the AMIDA13data.

Results: The deep learning system outperformed the traditional system over this homogeneous challenge dataset. However, we found sensitivity improved by combining the two systems. Specifically, atarate of 0.2 false detections per true positive, the traditional, deep learning, and hybrid systems' sensitivities were 38%, 48%, and 53%, respectively.

Conclusions: A combination of the two techniques yielded a hybrid system that out performed either system individually. This implies that information captured by the domain-specific features employed by traditional approach was not captured by the deep learning system.

J Pathol Inform. 2019 Sep 16;10:28.

Using High Resolution Android Smartphone Camera and Image Composite Editor to Create Whole Slide Images of Biopsy Specimens

Waqas Mahmud 1, Yahya Al-Ghamdi 1, Fatima Mir 1, Prih Rohra 1

Background: Due to small size of biopsy specimens it may be practically feasible to capture and merge images to create a single digital image of the entire tissue on the glass slide. While smartphone images of glass slides are restricted to informal use only, such images can offer an inexpensive alternative to slide scanning for low resource labs In low income areas. These images can also be used conveniently for slide sharing, consultation and learning. The objective of our study is to construct composite whole slide images of biopsy glass slides using android smartphone and compare them with light microscopy. Intra-observer diagnostic variability was used as the parameter to determine image quality.

Methods: 27 biopsy slides were selected from four organs including lungs, breast, prostate and colon. The selected cases included a broad range of common benign and malignant entities [Table 1]. Google pixel 2XL android phone with the following features was used to take images of the slides; Display: 1440 x 2880 pixels, 18:9 ratio, ~538 ppi density, Camera 12.2 MP, f/1.8, 27mm wide. Smartphone adapter was used to position the phone on a microscope with 30mm (10X) eye piece. Sequential Images were taken at low power (10X) with each image overlapping the preceding image. All the images were transferred to image composite editor (ICE), a Microsoft desktop application. The application stitched the images to produce a single merged digital image of the entire tissue (WSI).

Table 1.

Cases selected for use in the study and their concordance between light microscopy and smartphone generated digital images

Diagnostic entities Number of cases Concordance (yes) or major/minor discordance
Pathologist 1 Pathologist 2 Pathologist 3
Lungs
Adenocarcinoma 2 Yes Yes Yes
Squamous cell carcinoma 1 Yes Yes Yes
Poorly diff. carcinoma 1 Yes Yes Yes
Small cell lung carcinoma 1 Yes Yes Yes
Carcinoid 1 Minor (grade) Minor (grade) Yes
Squamous metaplasia 1 Yes Yes Yes
Bronchitis 1 Yes Yes Yes
Pneumonia 1 Yes Yes Yes
Breast
Invasive ductal carcinoma 2 Minor (grade) Minor (grade) Minor (grade)
Ductal carcinoma in situ 2 Yes Minor (grade) Minor (grade)
Usual ductal hyperplasia 2 Yes Yes Yes
Colon
Tubular adenoma without 1 Yes Yes Yes
High grade dysplasia
Tubular adenoma with 1 Yes Minor (grade) Yes
High grade dysplasia
-Invasive adenocarcinoma 1 Yes Yes Yes
Prostate
Benign prostate 5 Yes Yes Yes
Adenocarcinoma 4 Minor (grade) Yes Minor (grade)
Open in a new tab

Results: Adequate patient history was provided to three pathologists who independently looked at the digital images first and formulated a diagnosis. Later they reported the diagnosis by light microscopy. The diagnosis were compared between digital images and light microscopy. All three pathologist had minor discordance with their diagnosis in grading the tumor based on cytologic features, in 3/27 (11%), 3/27 (11%) and 4/27 (14%) cases. Notably, no major discordance, such as benign vs malignant or invasive vs noninvasive carcinoma, was observed.

Conclusion: Composite whole slide images of biopsy glass slides created by high resolution smartphone camera show high diagnostic concordance with light microscopy.

J Pathol Inform. 2019 Sep 16;10:28.

Deep Learning Based Method for Steatosis Quantification in Whole-slide Liver Histopathology Images

Mousumi Roy 1, Fusheng Wang 1, Jun Kong 2,3,4

Background: Steatosis quantification in liver histopathology images has high clinical significance. However, this process is challenging due to the large number of overlapped steatosis droplets with missing or weak boundaries. In this study, we propose a region-boundary integrated network for DeEp LearnINg stEATosis sEgmentation (DELINEATE). The resulting steatosis measures from our method present strong correlation with pathologist annotations, radiology readouts, and clinical data, suggesting its promising clinical deployment potential.

Technology: DELINEATE model can identify individual steatosis droplets and delineate boundaries of overlapping steatosis. As each whole slide tissue image has overwhelmingly large resolution, we partition each image into overlapping patches for analysis. The results from individual patches are seamlessly assembled with efficient spatial indexing-based matching.

Design: DELINEATE model consists of two sequential steps: a region extraction and boundary prediction model, followed by assembly of these two output predictions for final prediction map generation. The region extraction model is designed by stacking dilated convolution at the bottleneck of U-net model (dil-Unet). We use Holistically-Nested Network (HNN) as boundary detection model which learns hierarchically embedded multi-scale edge fields for object boundaries. A fully convolutional network (FCN-8s) is used to integrate the derived region and boundary information for final prediction result. Missing steatosis boundaries are recovered and assembled from adjacent image patches for whole-slide tissue analysis. The human subjects were approved by Emory University IRB.

Results: We use 100 whole slide microscopy images of human liver biopsies. Table 1 summarizes performances of DELINEATE, fully convolutional network (FCN) and DeepLab. Of all methods for comparison, DELINEATE model dil-Unet+HNN+FCN-8s achieves the best performance by both quantitative and qualitative results. The efficacy of DELINEATE is further verified by its strong correlations with liver tissue pathological grading and patient clinical data.

Table 1.

Summarizes performances of DELINEATE, fully convolutional network and DeepLab

Approach F1-score Object wise dice index Object wise Hausdorff distance
FCN 0.92±0.04 0.8338 3.8521
DeepLab V2 0.90±0.05 0.9083 5.3179
dil-Unet + HNN + FCN-8s 0.94±0.03 0.9492 3.4591
Open in a new tab

FCN: Fully convolutional network

Conclusions: The proposed DELINEATE network is based on joint use of the perceptual information from steatosis internal regions and boundaries for overlapped steatosis droplets, presenting a promising potential for accurate histology measurement for clinical use.

J Pathol Inform. 2019 Sep 16;10:28.

Establishing a Quality Control Process for Clinical CD8 Quantitative Image Analysis Testing

Lindsey Seigh Med 1, Douglas J Hartman 1, M S L Kathleen Cieply 1, Liron Pantanowitz 1

Background: In pathology all clinical laboratory testing should have quality control (QC) checks to ensure that reliable patient results are reported. Documentation of QC measures is important for compliance purposes. However, guidelines related to performing QC for quantitative image analysis are limited. Our aim was to develop a QC process for a laboratory developed semi- quantitative image analysis test used for clinical reporting.

Technology: Image analysis was performed using a laboratory developed algorithm to quantify CD8 immunohistochemical (IHC) staining in whole slide images. We used the Leica digital pathology platform comprising AT2 whole slide scanners and eSlide manager (Leica Biosystems Imaging, Inc., Vista, CA, USA) interfaced with our anatomical pathology laboratory information system (CoPath Plus, Cerner, MA, USA).

Methods: The image analysis laboratory employed three different QC measures [Table 1]. For daily testing a slide that contains positive/negative CD8 IHC controls is selected and run as the batch control. Weekly analysis of another IHC assay control ensures that no changes occur between patient cases. The precision control measured on the same slide each week was used to assess the entire digital pathology process. All QC results are recorded using Microsoft Excel and monitored regularly.

Table 1.

Three quality control parameters used for clinical CD8 quantitative image analysis testing

QC parameter QC Measure Purpose QCMethod Potential Problems Identified Example ExcelQC Result
Batch Control Check that IHC is working well Analyze any new patient slide from a batch of IHC IHC histology error & digital pathology system error graphic file with name JPI-10-28-g004.jpg
Control Slide QC Check that IHC controls are acceptable each week Analyze new IHC control slide each week IHC histology error, error with control tissue & digital pathology system error graphic file with name JPI-10-28-g005.jpg
Precision Control Check the precision of the digital pathology process each week Analyze the same IHC slide with same conditions each week Drift in either whole slide imaging or image analysis performance graphic file with name JPI-10-28-g006.jpg
Open in a new tab

Results: Determining which QC metrics to use with image analysis testing was challenging. Not only were there limited references available, but notable cellular changes (i.e. CD8 cell density) occurred on each deeper tissue section even with the same controls. Precision measurements allowed the lab to troubleshoot potential issues with the imaging process (e.g. image acquisition variation), while the weekly assay control was more helpful at detecting histology issues (e.g. stain artifact). These weekly QC measures provided a better evaluation of the entire test process compared to just batch QC measurements.

Conclusion: Although clinical laboratory accrediting agencies only require batch controls for digital image analysis, based on our experience also performing weekly precision and IHC assay QC measurements is recommended to better monitor the entire image analysis process from histology through to image computation. Displaying QC results graphically is advocated for quick visual acceptability prior to generating clinical reports and for regular medical director review.

J Pathol Inform. 2019 Sep 16;10:28.

Robust Epithelial Cell Staining in Breast Tissue for Light Sheet Microscopy

Paul D Simonson 1, Adam K Glaser 2, Suzanne M Dintzis 3, Jonathan T C Liu 2

Background: Light sheet microscopy is an imaging modality that can create three-dimensional images of specimens in short scan times while preserving tissue for additional ancillary studies. While thick tissue specimens can be stained using fluorescent dyes that readily penetrate tissue, specific antigen staining, usually performed using antibodies, has been less well demonstrated. Specific staining of epithelial cells can be key in identifying carcinoma, as in the case of infiltrating lobular or ductal breast carcinomas. As we are interested in the development of light sheet microscopy for breast surgical margin assessment and HER2 expression, we set out to develop epithelial cell antibody staining compatible with light sheet microscopy.

Methods: Fresh breast tissue was fixed in 10% formalin overnight, then incubated for 8 hours in blocking buffer containing 1% normal mouse serum (v/v), 1% BSA (w/v), and 0.3% Triton X-100. Tissues were next incubated for two days at room temperature with anti-EpCAM antibodies (Ber-EP4) that had been conjugated directly to DyLight 650, diluted in blocking buffer. Stained samples were then washed in blocking solution for 6 hours, then placed in a solution of TDE in water (n=1.46) overnight. Imaging was performed using a custom light sheet microscope. Similar procedures were followed for imaging dye-conjugated BSA and isotype antibody controls.

Results: Breast epithelial cell membranes were well stained and exhibited a signal to background ratio of >10x for tissue depths up to 1 mm [Figure 1]. Control staining was appropriately negative. Repeated trials with slightly varying staining times and incubation conditions demonstrated essentially equivalent results.

Figure 1.

Figure 1

Open in a new tab

image of breast tissue with ductal cells highlighted by Ber-EP4- Dylight 650 conjugate

Conclusions: The presented method results in robust staining of breast epithelial cells in light sheet microscopy. Of note, no antigen retrieval methods were required, and the staining worked with fixed tissue. While cytoplasmic cytokeratin staining is commonly used to identify epithelial cells, the use of Ber-EP4 likely presents a significant advantage in that antibody penetration through a fixed cell membrane is not required. As EpCam is expressed on a variety of epithelial cell types, it is likely that this protocol will be useful in a variety of tissue types. Further work will involve testing with solvent-based clearing techniques.

J Pathol Inform. 2019 Sep 16;10:28.

Quantitative Digital Image Analysis to Evaluate Podocalyxin as a Novel Immunohistochemical Biomarker of Placental Vascular Changes in Preeclampsia

Lauren Brilli Skvarca 1, W Tony Parks 2, Carlos Castro 3,4, Amber Luketich 3, Jeffrey L Fine 5, Janet M Catov 4, Carl A Hubel 4

Background: Preeclampsia increases a woman's risk of developing cardiovascular disease by two-fold within 5-15 years. Data suggest that impaired maternal vascular adaptation during placental development both contributes to multi-organ disease intrapartum and unmasks maternal cardiovascular risk postpartum. In the placenta, decidual vasculopathy (DV) is associated with preeclampsia and reflects impaired maternal vascular adaptation, possibly through endothelial injury and repair. DV is associated with increased risk of maternal cardiovascular disease after preeclampsia and increased risk of future adverse pregnancy outcomes even when identified in an uncomplicated pregnancy. Several diagnostic challenges limit consistent identification of DV by pathologists. Here we conducted a pilot study of podocalyxin (PODXL), a glycoprotein critical for endothelial barrier function, as a potential biomarker of DV and target of placental vascular disruption during preeclampsia.

Methods: Deidentified, formalin-fixed paraffin-embedded tissue from delivered placenta/membranes was obtained according to approved institutional policies for human tissue procurement (STUDY#18110123). We performed immunohistochemical staining with anti-PODXL antibody (Millipore-Sigma, St. Louis, MO, USA) and 3,3'-diaminobenzidine chromogen, followed by whole slide scanning at 40X magnification and image analysis using Aperio AT2 scanner and ImageScope software (Leica Biosystems, Buffalo Grove, IL, USA). Regions of interest (ROIs) were drawn encircling vascular segments in intermediate villi, terminal villi, or decidua. We used the ImageScope Positive Pixel algorithm to quantify staining intensity by calculating the sum of medium-positive and strong-positive pixels divided by length for each ROI (outputs “Np” + “Nsp” / “length”). Averages were compared between vessel groups by Student's T test.

Results: To optimize quantification, we analyzed PODXL endothelial staining in n=2 placentas from term, normotensive pregnancies. We compared staining intensity in intermediate villous vessels, terminal villous vessels, and decidual vessels (n=50 per vessel type). Intermediate villous vessels and decidual vessels exhibited significantly greater staining compared to terminal villous vessels [*p<0.005, Table 1, mean±standard deviation]. We quantified PODXL in decidual vessels that exhibited DV (n=14). Compared to normal decidual vascular segments, DV vessels showed significantly increased PODXL staining [† p<0.005, Table 1].

Table 1.

Podocalyxin staining quantification using image scope positive pixel count algorithm

Vascular compartment Intermediate villi Terminal villi Decidua: Normal vessels Decidua: DV vessels
ROI 50 50 50 14
Average ROI length (um) 462±264 121±42 548±444 342±182
Medium-positive (Np) + strong positive pixels (Nsp)/length (#/um) 27±19* 9±6 31±12* 51±15†
Open in a new tab

*P<0.005, mean±standard deviation, †P<0.005. ROI: Regions of interest

Conclusions: These data demonstrate feasibility of digital image analysis to quantify PODXL immunohistochemical staining in placental vasculature. Our preliminary findings suggest PODXL warrants further investigation as a novel DV biomarker.

J Pathol Inform. 2019 Sep 16;10:28.

A Comparative Study of Intraoperative Tele Neuropathology with Conventional Intraoperative Glass Slide Consultations

S U Baskota 1, C Wiley 1, L Pantanowitz 1

Background: At our institution the neuropathology division performs intraoperative consultations by both conventional (manual glass slide) and telepathology methods. The aim of this cross-sectional study was to compare the deferral and concordance rates for intraoperative diagnoses rendered using telepathology and without telepathology.

Technology: Aperio LV1 4-slide capacity scanners with an attached desktop console (Leica Biosystems, Vista, CA, USA) and GoToAssist (v4.5.0.1620, Boston, MA) were used for all intraoperative telepathology cases.

Design: A cross-sectional comparative study was conducted comparing 3 different hospitals (UPMC Mercy 25 cases, UPMC Shadyside 118 cases and UPMC Children's Hospital of Pittsburgh 22 cases; total 165 cases) where tele neuropathology was undertaken to UPMC Presbyterian (165 cases) where intraoperative neuropathology consultation was performed by conventional glass slide examination at a light microscope. Deferral and concordance rates were compared to final histopathological diagnoses. Concordance was divided into 6 categories: inadequate for diagnosis, no lesional tissue, wrong tumor type, wrong pathological process, correct category but different grade and exact diagnosis [Table 1 and Figure 1a]. Discordant cases were subcategorized into 4 different categories: neoplastic to non-neoplastic, non-neoplastic to neoplastic, inadequate for diagnosis of tumor, different type of tumor [Figure 1b–d].

Table 1.

Different categories of concordance with their rates

Concordance category Intraoperative telepathology cases Intraoperative manual cases Final diagnosis after telepathology Final diagnosis after manual
Correct category but variedgrade 94 53 23 20
Exact diagnosis 47 75 123 118
No lesional tissue 17 21 15 19
Wrong tumor 3 4 2 1
Inadequate for diagnosis 3 1 0 1
Wrong pathologic process 1 11 2 6
Total 165 165 165 165
Open in a new tab

Figure 1.

Figure 1

Open in a new tab

(a) Concordance rate telepathology versus no telepathology Discordance case example: (b)In intraoperative telepathology, it was called low-grade glioma (H & E stain, 20x magnification) (c) The same case was signed out as CD20 positive lymphoma in final histopathology (H & E stain, 20x magnification) (d) Discordant cases intraoperative telepathology versus manual to final histopathology

Results: The deferral rate for intraoperative telepathology was 26% and conventional glass slide 24.24% (p-value 0.58). The concordance rate for telepathology was 93.94%, which is slightly higher than 89.09% of conventional glass slides (p-value 0.047).

Conclusion: Intraoperative telepathology interpretation of neuropathology frozen specimens at our institution is as effective as conventional glass slide interpretation. The complexity of cases in the randomized groups and interobserver variation among neuropathologists, who are all familiar with telepathology technology at our medical center, can be a limitation and introduce bias into this study.

J Pathol Inform. 2019 Sep 16;10:28.

Pathology Informatics Research Using Publicly Available Databases: A Case Study

Kingsley Ebare 1, Yonah C Ziemba 2, Tarush Kothari 2

Background: Pathology Informaticists are uniquely positioned for a leadership role in cancer research given their combined expertise in cancer pathology and data analysis. In particular, publicly available databases are a valuable and under-utilized resource. Surveillance, Epidemiology and End Result (SEER) is a public database of special interest to pathologists because of the available pathology information, including diagnosis, diagnostic modality, metastasis, tumor focality, grade, stage and survival. To demonstrate the utility of SEER for pathology research, we present an analysis of tumor grade as a prognostic factor for papillary urothelial carcinoma of the ureter and renal pelvis.

Technology: Data was downloaded from the SEER portal, publicly available at https://seer.cancer.gov. Analysis was conducted in SAS 9.3.

Methods: All records of histologically confirmed papillary urothelial carcinoma from 2010 to 2015 were considered. Records were excluded if they lacked information regarding WHO/ISUP grading, disease stage or active follow-up. Mortality of high grade versus low grade carcinoma was compared using Cox proportional hazard model in univariate and multivariate analyses.

Results: A total of 1352 cases were identified, and 927 were excluded because they lacked grade, stage or follow-up. The remaining 425 cases comprised our study cohort. 55 (13%) of these were low grade and 370 (87%) were high grade. Univariate analysis showed increased mortality in high grade carcinoma compared to low grade carcinoma as depicted in the Figure 1 (HR=2.40; 95% CI =1.14- 6.16). However, when multivariate analysis was performed to control for age at diagnosis and disease grade, there was no significant increase (HR=1.408; CI=0.66–3.66). However, when stratifying by stage, increased mortality in high grade was seen only in stages 3 and 4 (HR=3.55; CI=1.32-9.52 and HR=16.37; CI=6.33-42.38 respectively). There was no statistically significant difference in stage 1 and 2 (HR=1.41; CI = 0.61-3.27 and HR=1.57; CI = 0.40-6.07 respectively).

Figure 1.

Figure 1

Open in a new tab

Cancer specific curve of urothelial carcinoma of renal pelvis and ureter stratified by grade

Conclusion: Publicly available databases are goldmine of opportunity for pathology informaticists. They can be used to gain prognostic and diagnostic insights, and to generate hypotheses for further research.

J Pathol Inform. 2019 Sep 16;10:28.

Reference Interval Validation for Intact Parathyroid Hormone with Raw and Resampled Data

J Abel 1, A Fnu 1, D Bowman 2, E S Pearlman 3

Background: In January, 2017, the vendor of our chemistry equipment (Ortho Clinical Diagnostics [OCD]; Rochester, NY, USA) informed us of a positive bias in measurement of iPTH. No change in the manufacturer's reference interval (MSRI) [7.5-53.5 pg/mL) was suggested and no indication was given of when the problem started. Our lab collected all PTH results from calendar years 2014-2016 for re-evaluation of MSRI. In February, 2018 we were notified that the vendor that the MSRI could no longer be supported as the upper limit was too low.

Technology: iPTH assay used the Vitros 5600 analyzer (OCD; Rochester, NY). Calculations were performed using Peak Fit and Table Curve-2D software (Systat, Chicago, IL). Resampling was performed using the R package “GSM” (Vienna, Austria).

Design: The data containing both “normal” and pathologic results was modelled as a mixture of distributions, typically gamma. The results were binned in intervals of 3.5 since data was left censored at 3.4 pg/mL. As patients with results outside the MSRI were tested more frequently, multiple patient values were replaced with a mean. To delete the influence of large but infrequently occurring values the data was resampled. LOESS smoothing and curve fitting used Peak Fit software. Based on prior knowledge the upper endpoint was chosen to maximize sensitivity for detection of disease. All data greater than the putative upper limit was deleted and the remaining data points were re-binned, smoothed using splines and a lower limit chosen to encompass 95% of the AUC using Table-Curve 2D. Results using both raw data and data in which multiple patient measurements had been averaged were compared.

Results: Results are shown in the accompanying Table 1.

Table 1.

Computed reference intervals (95% AUC) using gamma or generalized error mixture models for raw, averaged, resampled, and/or original data with associated r2 values

Data type Resampled peak count Distribution r2 95% AUC interval (pg/mL) Maxsensitivity point (pg/mL)
Raw N/A Gamma 0.982 8-118.5 123.5
Average, resampled 10 Generalized Error 0.994 8-80.5 82.1
Average, resampled 100 Gamma 0.957 8-110 117.7
Raw, resampled 10 Gamma 0.957 8-107 109.9
Raw, resampled 100 Gamma 0.973 8-118.5 120.2
Open in a new tab

N/A: Not available, AUC: Area under the curve

Conclusion: Our RI was 57% larger than the MSRI. when using the resampled, averaged data. The upper end point was indeed considerably higher. Manufacturing problems were presumably the major cause but our patient population being 90% male and 32% African- American may have been quite different from the population used to derive the MSRI.

J Pathol Inform. 2019 Sep 16;10:28.

Introducing the Pathology Informatics Essentials for Residents App for Pathology Residents

Robin Dietz 1, Liron Pantanowitz 1

Background: Pathology Informatics Essentials for Residents (PIER) is now an approved education curriculum for pathology residents. PIER contains multiple sections including content about informatics activities, practical exercises, resources, and a sign-off mechanism by an attending pathologist. Currently PIER is structured in an interactive 85-page Portable Document Format (PDF). We sought to present the PIER curriculum as an iPhone app for easier use and the potential for adding novel features.

Technology: The app was designed with Xcode Version 10.1 using Swift programming language. It uses Firebase, Google's mobile platform, and Cloud Firestore, a horizontally-scaling NoSQL cloud- based database. The database contains higher-level “collections” that contain multiple “documents.” The documents each contain up to 1 MB of JSON-structured data [Figure 1a]. Unlike traditional SQL databases, this allows saving of unique fields in each document.

Figure 1.

Figure 1

Open in a new tab

(a) The PIER app uses a NoSQL database, a horizontally-scaling database that allows saving of unique field types, using Google Firebase. (b) Screenshot of the PIER app welcome page, showing that the structure of the educational content is similar to the PDF version

Design: The PIER app is designed to allow attending pathologists to review and sign-off on the progress of their trainees, as well as for pathology residents to easily enter data [Figure 1b]. The app could be used for real-time monitoring of resident's progress, sharing informatics findings between residents, and as a means of dispersing informatics news to end users.

Results: The app structure mimics the layout of the PIER content from the PDF version. It permits data to be saved and sent for review by an attending pathologist. It offers a quick way for pathology residents to record their journey through pathology informatics at their institution and allows them to record data in a way not permissible by PDF. At the same time, it maintains current functionality of the PIER PDF document.

Conclusions: Offering an app to easily navigate the PIER curriculum has potential for enhancing pathology resident's informatics educational experience. Additional benefits include real-time monitoring of trainee progress and access to up to date resources. Once launched this app will be offered to trainees at our institution. We are also currently working with the PIER Leadership Committee about more widespread implementation.

J Pathol Inform. 2019 Sep 16;10:28.

Clinical Correlation to a Putative Reference Interval

Alnoor Fnu 1, Jacob Abel 1, Dale Bowman 2, Eugene Pearlman 3

Background: In February, 2018 the lab at the Memphis VAMC received a communication form its' vendor of chemistry equipment (OCD; Rochester, NY) indicating that its' manufacturer's suggested reference interval (MSRI) {7.5-53.5 pg/mL] could no longer be supported as its' upper limit was too low. This followed an earlier warning in January, 2017. Initial workup of patient data from the 2014-2016 calendar years (N=7915) suggested a reference interval that was not much different at its' lower end but was possibly as high as 100 pg/mL at its upper end. We were motivated to examine the distribution of clinical diagnosis at different segments of the data range.

Technology: Ten patient records in four different segments of the data range assuming a preliminary RI of 7.5-100 pg/mL were selected by convenience sampling for review [<7.5 pg/ml, 8-11 pg/ml, 90-100 pg/ml, 110-120 pg/ml.]. Since patient record numbers had been recorded, the patients' VA electronic medical records were readily accessed.

Methods: Using hospital electronic medical records, relevant patient data in each PTH concentration range was collected.

Results: In the ten patients in the lowest concentration range there was no diagnosis of primary hypoparathyroidism. The most common diagnosis was malignancy (n=6) including lung (3), prostate (1), bladder (1) and lymphoma (1). Hypercalcemia was identified as a problem in 5 of these patients. In the 8-11 pg/mL group, malignancy was identified in 8 patients including multiple myeloma (3), lung (2), prostate (2) , kidney (1)with hypercalcemia noted in 4 patients. With concentrations between 90-100 pg/ml, renal disease is most common (n=6) and 3 patients were noted to be Vitamin-D deficient. In the highest concentration range there are 7 patients (70%) with a diagnosis of renal disease and 3 with primary hyperparathyroidism.

Conclusion: The multiplicity of diagnoses involving the calcium-PTH feedback loop in the 8-11 pg/mL. group suggest that a low end concentration of 7.5 pg/mL may be too low. This may also suggest that an upper limit of 100 pg/mL. is too high but a threshold of 110 pg/mL. may be useful in distinguishing secondary from primary hyperparathyroidism.

J Pathol Inform. 2019 Sep 16;10:28.

Explainable Artificial Intelligence in Pathology: A Multifaceted Framework to Guide Development and Evaluation

Thomas J Gniadek 1, Jason Kang 1, Jacob Krive 2,3,4

Background: For decades, artificial intelligence (AI) algorithms were “black box” processes devoid of clear rationale for their output or robust confidence assessments, limiting AI's application and utility in medicine. Explainable artificial intelligence (XAI) focuses on combining “explanations” with core predictive AI algorithm output(s). XAI may reshape AI integration into medicine, especially for data-driven interpretive tasks in Pathology.

Methods: We assess primary requirements for successful XAI integration into the practice of Pathology. A framework is presented for classifying XAI algorithms based on how they integrate into the clinical workflow and how their validation could differ from the validation of traditional AI algorithms.

Results: XAI can report empiric assessments of predictive confidence or the degree of association between a case and historical patient populations or established disease mechanisms [Table 1]. If associations are presented, clinicians can independently evaluate the association validity and the core output that depends on it. XAI training and validation can consider not only the validity of the core algorithm output, but also actions taken by clinicians as a result of the core output and explanation provided. Results can be displayed textually or graphically, with real-time input from clinicians regarding their confidence in the results and explanations provided.

Table 1.

Examples of explainable artificial intelligence (XAI) type and content

XAI explanation type XAI explanation content
Empiric Statistical confidence based on historical sample data
Population associative Association with historical population data
Mechanism associative Association with known pathologic mechanism (s)
Open in a new tab

Conclusions: XAI is heterogeneous in design and workflow. Algorithm classification will be needed when evaluating XAI and studying its efficacy. Usability and output type (textual vs. graphical) may have profound effects on outcomes. XAI offers immense promise to Pathology, with the goal of supporting rather than replacing the Pathologist.

J Pathol Inform. 2019 Sep 16;10:28.

Deep Learning to Predict Histopathology Findings from Gene Expression in a Mouse Collagen Antibody-Induced Arthritis Model

Tao Fang 1, Jitao David Zhang 1, Pierre Maliver 1, Alexia Phedonos 1, Claudia Bossen 1, Timo Schwandt 1, Matthias Wittwer 1, Annie Moisan 1, Virginie Sandrin 2, Mark D Robinson 3, Klas Hatje 1

Background: Assessing histopathology in animal models is critical to characterize the safety profile of newly developed drugs in vivo. Recent developments in next-generation sequencing and machine learning allow for new approaches to complement histopathology findings with insights of molecular and gene regulatory mechanisms. We investigated to which extent histopathology findings can be predicted from gene expression data alone using state-of-the-art machine learning methods.

Methods: A deep neuronal network and a support vector machine model were trained using publicly available data from the Open TG-GATEs (Toxicogenomics Project-Genomics Assisted Toxicity Evaluation System) database, which is one of the most comprehensive toxicogenomics databases and consists of drug-induced gene expression profiles from in vivo experiments and corresponding toxicological information including histopathology findings. Both machine learning models were applied to transcriptomics data derived from liver and kidney samples of a mouse collagen antibody-induced arthritis model and compared to histopathology findings from the same samples. The experimental design allowed to investigate samples with and without induction of inflammation as well as treatment with three anti- inflammatory drugs. In addition, the models were evaluated on liver and kidney transcriptomics data from naive rats treated with the same three compounds.

Results: The evaluation of the predictions revealed that the deep neural network results are more robust compared to the classical, support vector machine-based approach. Furthermore, the deep-learning approach provides fewer false positive predictions as a comparison to histopathology findings from the same experiments revealed.

Conclusions: Toxicogenomics data is a valuable source of information to investigate drug safety preclinically. These data already complement histopathology findings, but with the growing amount of reference data and new machine learning methods, they might allow for robust prediction of pathology findings from gene expression alone.

J Pathol Inform. 2019 Sep 16;10:28.

Distributed Automated Processing of Clinical Genomic Cancer Panels Leveraging Cloud Infrastructure

Caylin Hickey 1, Cody Bumgardner 1, Kannabiran Nandakumar 1

Background: With the recent development of next-generation sequencing (NGS),[1] more institutions are looking to leverage genomic sequencing for both academic research purposes as well as clinical cancer diagnostic assistance. In clinical applications, this can take the form of pipelines which translate some or all of the process by which raw base call images are used to generate annotated variant caller fi for bioinformaticists to provide diagnostic to the medical community. For this work, we describe the control framework used to manage this workflow for our clinical operations.

Technology: The complex, multifaceted computational requirements ofNGSrequirelarge amountsof processing power while also working on tightly controlled, mostly contained processes. As such, this system lends itself well to cloud computing infrastructure in which resources can be allocated as required. For our work, this has taken the form of both a private cloud, backed by OpenStack[2] and Ceph,[3] as well as the public cloud (Amazon Web Services) using encrypted architecture.[4] Additionally, as the work must be reproducible, we leverage Docker application containerization to precisely control the processing environment. Finally, we also leverage the BagIt specification for archiving data developed by the Library of Congress.[5]

Methods: Our framework is divided into two cooperative systems: Docker containers housing the processing tools (Illumina[6] BCL to FASTQ conversion (preprocessing,) sequence alignment, variant calling, variant annotation, and quality control metrics generation) and the agent- based control framework which archives the data at each step, delivers said data to the nextstage (managing the creation/deletion of the processing agent if required,)andtracking progress.[7] The system allows us to scale our processing to the NGS work available.

Results: Required time to process a two-lane run on an Illumina sequence, which equates to up to 16 samples of a 16 gene cancer panel, to under 24 hours for annotated results.

Conclusions: This framework has increased the scalability of our clinical and research genomic process- ing operations. Additionally, we now have the ability to scale to additional types of cancer panels by simply adding new Docker containers.

References

  • 1.Behjati S, Tarpey PS. What is next generation sequencing? Arch Dis Child Educ Pract Ed. 2013;98:236–8. doi: 10.1136/archdischild-2013-304340. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Sefraoui O, Aissaoui M, Eleuldj M. OpenStack: Toward an open-source solution for cloud computing. Int J Comput Appl. 2012;55:38–42. [Google Scholar]
  • 3.Weil SA, Brandt SA, Miller EL, Long DD, Maltzahn C. Proceedings of the 7th Symposium on Operating Systems Design and Implementation. USENIX Association; 2006. Ceph: A scalable, high- performance distributed fi system; pp. 307–20. [Google Scholar]
  • 4.Fusaro VA, Patil P, Gafni E, Wall DP, Tonellato PJ. Biomedical cloud computing with amazon web services. PLoS Comput Biol. 2011;7:e1002147. doi: 10.1371/journal.pcbi.1002147. [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5.Kunze J, Littman J, Madden E, Scancella J, Adams C. The Bag It File Packaging Format (V1. 0) Tech Rep. 2018 [Google Scholar]
  • 6.Illumina Sequencing. [Last accessed on 2019 Mar 10]. Available from: https://www.illumina.com/techniques/sequencing.html .
  • 7.Bumgardner VC, Marek VW, Hickey CD, Nandakumar K. 2016 IEEE 18th International Conference on e-Health Networking, Applications and Services (Healthcom) IEEE; 2016. Constellation: A secure self- optimizing framework for genomic processing; pp. 1–6. [Google Scholar]
J Pathol Inform. 2019 Sep 16;10:28.

Real Time Breast Histology Image Classification with a Mobile Phone

Kenji Ikemura 1

Background: Deep learning, specifically convolutional neural network, has made a breakthrough in the complex task of computer image recognition. In this study, depthwise separable convolutional neural network (DS-CNN) is used in classifying breast cancer histology images on the computer and on the mobile phone in real time. The study propose that DS-CNN can be applied for histological image analysis and its network can be transferred to a commercially available mobile phone for real-time histological image analysis captured through the mobile phone camera.

Methods: This study utilizes the DS-CNN on breast cancer images (training and test datasets) downloaded from publicly available repository: https://rdm.inesctec.pt/dataset/nis-2017-003. Training set images are augmented by rotation and mirroring the images. DS-CNN is trained to match the goal to classify breast histology images between 4 categories: i) normal, ii) benign, iii) carcinoma in situ, and iv) invasive carcinoma. Finally, the trained DS-CNN is deployed on to a mobile phone to classify the images captured through the mobile phone camera. The output on the mobile phone screen are the real-time image from the camera and its probability of it being one of the 4 categories (in highest to lowest rank order). Accuracy of DS-CNN is assessed by whether its output class with highest confidence matches the true class in the test dataset.

Results: The trained DS-CNN accuracy on the computer reached as high as 86% in 4 class classification. On the mobile phone, accuracy reached 69% in 4 class classification and 83% in 2 class classification (normal or benign vs. in situ or invasive). Training time took less than 30 min on 1.4 GHz Intel Core i5 dual-core CPU. Latency for evaluation on mobile phone was less than 1 second.

Conclusion: DS-CNN is a fast and efficient neural network architecture that can learn to distinguish histological images even with limited sample size and computational power. The architecture can be deployed onto a mobile phone and maintain relatively good accuracy through the phone camera. It is likely that the accuracy can be increased by expanding the dataset.

J Pathol Inform. 2019 Sep 16;10:28.

Sustained Benefits of Rules-Based Reflex Testing in Disease Association Testing

Jason Kang 1, John Cavataio 1, M T Christina Sutherland 1, Thomas Gniadek 1

Background: Modern laboratory testing often requires adherence to reflex test pathways. Multiple contingent, reflex test pathway options can exist in the same workup, which further complicates definitive interpretation and/or diagnosis. Laboratory management of such test pathways can be aided by rules logic. Our institution's Laboratory Information Systems (LIS) group leverages an integrated LIS in order to manage an in-house, rules-based diagnostic reflex test cascade for celiac disease.

Technology: The test cascade consists of fifteen rules that hinge on age and immunoglobulin A (IgA) level. The rules' logic incorporates patient age and IgA level in order to stratify patients into groups based on age (inclusive of the “to” and “from” age values) and accepted IgA level upper and lower limits. Based on the defined rules and conditions associated with these two values in each group, additional tests are performed in order to facilitate formal pathologist interpretation.

Methods: The current complete test cascade rules and logic were implemented during fiscal year (FY) 2015. The initial ordered test is an IgA level quantification. Based on this result and the patient's age, testing is reflexed to tests from the following options: Tissue Transglutaminase Antibody, IgA (TTGA); Tissue Transglutaminase Antibody, IgG (TTGG); Deamidated/Gliadin Antibody, IgA (DGLA); Deamidated/Gliadin Antibody, IgG (DGLG); Human Leukocyte Antigen (HLA) DQA and DQB Typing. Testing may be reflexed to multiple tests at once, depending on the applicable rules logic and conditions. Individual test results with pathologist interpretation are entered in the LIS.

Results: Formal, in-house pathologist interpretation is enabled in 100% of cases without the need for additional reflex test orders. Molecular laboratory test volume has sustained a mean increase of 160% in relevant HLA test volume (DQA and DQB antigen typing) since FY2015.

Conclusions: Rules-based test cascades improved access to in-house pathologist interpretation and increased billable in-house test volumes. More broadly, our institution's experience demonstrates that rules-based diagnostic reflex test cascades can be deployed in support of disease screening and diagnosis.

J Pathol Inform. 2019 Sep 16;10:28.

Anatomical Pathology Laboratory Augmentation Using Three-dimensional Printing

Rufei Lu 1, Zach Webb 1, Christopher William 1

Content: Although computerized 3-dimensional visualization has revolutionized structural engineering design, these models were still largely limited to 2D screen and greatly hampering practicality of the 3D models. The recent popularization of 3D printing technologies and open- sourced software have not only simplified the fabrication workflow, but also dramatically reduced the cost of custom designs. Despite much effort of standardizing all surgical pathology laboratories procedures, there are still countless unique technical challenges and instrumentation needs encountered by each individual laboratory. Here, we reported a cost- effectively workflow to overcome many of these specific laboratory needs by customizing surgical pathology tools using 3D printing.

Technology: The initial measurements of customized parts and laboratory instrumentation dimensions were ascertained using standard rulers and engineering fractional digital caliper. Google SketchUp Pro (version 17.3.116) was used for the model design and generation of stereolithography files (.stl). Model slicing and translation to G-code were accomplished using Ultimaker Cura (3.1.1). 3D models were printed using Monoprice Maker Ultimate 3D Printer - MK11 DirectDrive Extruder / 24V Power System. All models were created using 1.75 mm diameter polyactic acid (PLA) filament.

Design: Generic workflow for the model design and fabrication was accomplished with mostly open-sourced software as shown in Figure 1. The 3D models were initially created using Google Sketchup Pro due to its adaptability, intuitive user interface, and copyright protection. After the 3D models have been tested in a computerized theoretical environment, the files were converted to .stl files. Finally, slicing and G-code generation were accomplished using open- sourced Ultimaker Cura. The prototypes were created using PLA filament at 0.5 mm layer resolution and 3x outer shell setting. After prototypes have been tested in practice by laboratory technicians and residents, feedback was iteratively incorporated into original design and version controlled. The final products were printed using 0.1 mm resolution and 5x outer shell setting.

Figure 1.

Figure 1

Open in a new tab

Generic workflow for the model design and fabrication

Results: Using our workflow, we successfully created multiple customized tools that requested by surgical pathology and cytology technicians, residents, and faculties. Of note, we designed both single lamella and multi-lamellae slicing tools that not only created consistent sections accurately and efficiently, but also improved the safety of the resident who are grossing the specimen. We also created several customized mobile conical tube holders for cytology technologists, who constantly struggle with limited space during on-site adequacy evaluations. Additionally, we were able to create an exact replica of broken microtome part using PLA filament at 0.1 mm layer resolution and 5x outer shell setting. The part was successfully replaced into the broken microtome and potentially saved the department replacement cost.

Conclusion: The customized laboratory tools can be created in a cost-effective and timely manner. The durable 3D printed tools were able to accommodate the specific laboratory needs, expand existing workflow, and improve safety.

J Pathol Inform. 2019 Sep 16;10:28.

The Use of Machine Vision Cameras in Photomicroscopy, Hidden Treasures and the Cutting Edge

Mark Luquette 1

Background: Piezo shift digital cameras have been leaders in digital microscopy for 2 decades. Machine vision cameras are available with user friendly software that can compete with piezo cameras for near a tenth of the cost. This report studies the performance and ease of use of 4 static chip machine vision cameras and compares the results to a piezo shift camera (Olympus DP72).

Technology: (camera / type / pixels / pixel size μ / chip diagonal mm / source):

  1. Olympus DP72 / pixel shift / 4140 x 3096 / 6.54 / 11 / olympus-lifescience.com

  2. Imaging source DFK 33GP5000e / static / 2592 x 2048 / 4.8 / 16 / theimagingsource.com

  3. Imaging source DFK MKU130-10x22 / static-back lit / 4128 x 3096 / 1.4 / 7.2 / “

  4. Basla acA1920-40uc / static / 1920 x 1200 /5.86 / 13.35 / microvisioneer.com

  5. Teledyne Dalsa Genie TS-C4096 / static / 4096 x 3072 / 6 / 31 / teledynedalsa.com Dove tail connectors, RafCamera, Novopolotsk, Vitebsk, Belarus.

Modular optics, Thorlabs, Newton, NJ, USA.

Methods: Images of the same field were taken with the same objective, and compared based on what digital zoom point pixilation occurred. The piezo camera used a stock coupler- trinocular head. Machine vision cameras were coupled to the dual observer port using a dove tail connector.

Results: Results of the comparison study are seen in Table 1.

Table 1.

Camera comparison results

Camera Image (in) Pixelate at (%) Size (MB) Field of view µ Focal length (mm)
DFK 33GP5000e 36×28 100-200 11.4 890 125
Genie TS-C4096 56×42 100-200 36.9 880 250
DP72 38×29 66.7-100 38.1 712 -
Open in a new tab

Conclusion:

  1. A relatively inexpensive machine vision camera with a single lens coupler can produce images that compete with a piezo camera.

  2. Intuitive software is available.

  3. Back lighting tiny pixels did not perform.

  4. The Dalsa camera software was too complex.

J Pathol Inform. 2019 Sep 16;10:28.

The Current Landscape of Clinical Informatics Education

Emilio Madrigal 1, Anand S Dighe 1, Jason Baron 1, William J Lane 2, Danielle Kurant 2, W Stephan Black-Schaffer 1, Veronica E Klepeis 1

Background: The growth and complexity of data generated by modern pathology laboratories have increased demand for a specialized workforce well-versed in informatics. To meet this need, board certification in clinical informatics (CI) has been offered to eligible physicians in the United States (US) since 2013; formalized CI education through Accreditation Council for Graduate Medical Education (ACGME)-accredited two-year fellowship programs was introduced in 2014. We assessed the state of CI education in the US, with a focus on pathology.

Methods: Pathology (anatomic and clinical) residency programs were identified using the search tool provided by ACGME. We also extracted details about all CI fellowship programs (CIFPs) in the US including location, administering residency review committee (RRC), accreditation date, and approved positions. The American Medical Informatics Association's website was queried to identify the number of board-certified CI subspecialists. A choropleth map was generated to visualize findings.

Results: A total of 143 ACGME-accredited pathology residency programs were identified in 42 states. Among these programs, 32 CIFPs are approved for 118 positions across 18 states. Eleven states have only one CIFP, with the remaining 7 sponsoring two or more [Figure 1]. Of the 9 potential RRCs within the ACGME, only 6 currently administer CIFPs: internal medicine (n=13), pathology (n=7), pediatrics (n=5), emergency medicine (n=4), family medicine (n=2), and anesthesiology (n=1). Of note, over the past 2 years, 5 of the 8 most recently approved CIFPs (63%) are based in pathology. As of 2017, there are 1,690 board-certified CI diplomates, of which 6.3% (n=107) were certified by the American Board of Pathology (ABP).

Figure 1.

Figure 1

Open in a new tab

Accreditation Council for Graduate Medical Education accredited clinical informatics fellowship programs

Conclusion: Over the last half-decade, new ACGME-accredited CIFPs have provided a path to CI certification. More recently, the ABP has provided further impetus for pathologists to become certified by approving simultaneous completion of another one-year fellowship (with focus on CI) over the same 2 years as the CI fellowship. Although there are many CIFPs across the US, some regions are poorly represented, namely the southeast, which comprises close to one- quarter of pathology residency programs but is home to only 9% of CIFPs. The field of informatics is evolving, and new CIFPs are necessary to attract and educate the next generation of informaticians. Providing certification in CI recognizes that expertise and fosters the growth of the discipline.

J Pathol Inform. 2019 Sep 16;10:28.

Automatic Metadata Extraction from Tissue Slide Label Image

Yao Nie 1, Faranak Aghaei 1, Tomoyo Sasaki 2, Dennis Wang 2, Yifei Zhu 2

Background: Existing staining instruments allow users to enter important slide metadata to be printed on slide labels. However, additional manual data entry is usually required to enable metadata management in a system that cannot be linked to staining instruments directly. To reduce duplicated manual data entry, we developed and experimented with a software tool with graphical user interface to automatically extract metadata from slide label images, aiming to provide a cost-effective solution to improve data management in laboratory operations.

Methods: Two types of slides stained by Roche BenchMark Ultra and Dako Autostainer instrument were scanned by Ventana iScan HT scanner, and the label images were extracted from the scanned files. The type of each label was automatically identified using a convolutional neural network based image classification algorithm. Then an optical character recognition algorithm that was trained for each instrument was applied to identify the texts in each label image. The extracted metadata were exported to a csv file. Two processing workflows were built to provide flexible usability. Specifically, Individual Screening workflow allowed users to check and edit the extracted metadata one by one through user interface, and Batch Processing workflow processed all labels together and users corrected the results in the csv file in the end. Finally, the corrected csv file was uploaded to a tissue slide image management system.

Results: The tool was tested over 277 tissue slide labels with different challenges including label damage, rotation, stain over texts, etc. Label type identification achieved 100% accuracy. Individual Screening took one hour and achieved perfect accuracy. Batch Processing took one minute with an error rate of 15%. The corrected csv file was successfully uploaded and the metadata was shown correctly for each tissue slide image in the slide image management system.

Conclusions: The developed tool can automatically identify the label type with high accuracy and extract the metadata with two processing workflows. Both workflows were more efficient than manual data entry, which demonstrated the feasibility and effectiveness of implementing such cost-effective tool to improve laboratory operations. Further improvement can be achieved by a dedicated quality control interface to be used at the end of Batch Processing.

J Pathol Inform. 2019 Sep 16;10:28.

Using Open-Source Software for Education – Creating a Basic Whole Slides Images Collection in Two Weeks

Andrey Prilutskiy 1

Background: Using proprietary whole slide image viewer software from scanner manufacturer for teaching slides collection often imposes various restrictions in portability, and compatibility due to operating system restriction and limited file format support. We attempted to create a scalable, portable, cross-platform basic whole slide image teaching collection from the scratch, using open-source viewing software QuPath and open Pathology encyclopedia Librepathology.org.

Technology: We used the open-source free software QuPath 0.2.0 as the main viewing tool. Librepathology.org was used for entities selection. MindNode 6.0.1 iOS application used for building basic entities mind maps for the visualization of progress.

Methods: Most basic entities in each organ system have been chosen from Librepathology.org website. The list was then imported to MindNode application to create a mind map for visualization and progress tracking. Through Cerner CoPath search covering the most recent two years in our institution's on-site slide archive, representative slides for each entity were pulled for scanning. We attempted to include a least 2 representative cases for each entity. Ventana iScan Coreo scanner was used for scanning slides. Resulted TIFF whole slide image files were then organized to folders based on organ system and images in each folder then imported to separate QuPath projects [Figure 1].

Figure 1.

Figure 1

Open in a new tab

quPath software with the endocrine pathology collection

Results: Over 230 slides were pulled from the slide archive and scanned in two weeks' period and organized into a slide collection. Resulting whole slide images collection covers basic entities in gastrointestinal, genitourinary, gynecological, breast, endocrine, neuro-, and thoracic pathology. The collection can be used on department computers as well as on personal devices after de-identification of whole slide scans.

Conclusions: Due to cross-platform compatibility with Windows, MacOS, and Linux, QuPath software can be an excellent tool for creating a scalable and portable slide collection that can be used for junior residents and students' education. Multiple image formats support allows combining multiple sources of teaching slides. Mind maps can be used for tracking the process and visualizing the collection. Using open source software can improve access to virtual teaching sets among trainees and students.

J Pathol Inform. 2019 Sep 16;10:28.

Implementation of a Clinical Pharmacogenomics Workflow with Integrated Clinical Decision Support

Rebecca A Pulk 1, Nathanial Price 2, Fatma Issa 2, Dave Ferguson 3, Charles Torre Jr 2, Wade L Schulz 4,5

Background: Inherited differences in enzymes, transporters and immune markers can lead to inter-patient differences in medication response and safety. Routine clinical use of pharmacogenomics, the customization of medication dosing through the application of genetic information, has been hampered by the traditional reporting of these results as static reports along with a lack of standardized, structured data elements for this information within electronic health records. However, new approaches to genomic data management necessitate careful attention to clinical and regulatory requirements for laboratory reporting.

Technology: Health systems with robust pharmacogenomic practices have traditionally built custom electronic health record and laboratory information system data elements to support pharmacogenomic data management and clinical decision support. But new approaches to integrate real-time clinical decision support, including through web service interfaces, are emerging. We integrated a commercial pharmacogenomic clinical decision support platform, ActX (Seattle, WA), with our Epic electronic health record (Verona, WI) to implement such an approach.

Methods: We created an order with reflex testing to support billing for specific clinical indications while using a panel-based genotyping chip for evaluation. Existing ActX alerts were customized to include only medication-gene associations for indications meeting evidence standards. Site-specific language was deployed for active mediation alerts. Work with the developer has allowed for passive reporting of medication-gene interactions, a site-specific list of medications tested for, and a mechanism for reporting results based upon provider preferences and regulatory requirements.

Results: Our first wave deployment of this platform has allowed us to report on 29 medication-gene pairs associated with evidence-based guidelines from the Clinical Pharmacogenomics Implementation Consortium in a provider-friendly manner. Cloud-based storage of all generated genetic data allows for real-time clinical reporting with up-to-date information, while mirrored local variant data can be used for investigative research.

Conclusion: By implementing a vendor clinical decision support system with our existing electronic health record, we were able to bypass many bottlenecks associated with the clinical and technical implementation of a pharmacogenomics platform. However, additional customization was necessary to meet local provider preferences and regulatory requirements for laboratory testing and reporting.

J Pathol Inform. 2019 Sep 16;10:28.

Docker Container-based High-Performance Computing Environment for Next Generation Sequencing Data Analysis

Somak Roy 1, Marina N Nikiforova 1, Yuri E Nikiforov 1, Robert L Ferris 2, Aatur D Singhi 1

Background: Containers are a standard unit of software that provide isolated user space for packaging of applications and their dependencies and deploying across a wide array of computing environments. Containers have been predominantly used for deploying web application and API services. However, use of containers for HPC-like environments is currently very limited. This study aimed to use the Docker (Docker Inc. San Francisco, CA, US) framework to create an HPC environment for enabling NGS data analysis.

Methods: The compute cluster consisted of 3 Dell servers (Dell Inc., Round Rock, TX, US), each running Docker Engine CE (v18.09.1) in an Ubuntu server v18.04. This 3-node cluster was enabled using Docker Swarm over a 1Gbps network. Individual Docker images were built using Alpine Linux v3.8 base image with python v3.6 runtime. Persistent storage for containers was provisioned from a separate storage server over a 40Gbps network. Figure 1 provides the architectural details of the setup.

Figure 1.

Figure 1

Open in a new tab

Docker container-based high performance computing setup

Results: The swarm cluster had a total compute resource of 110 cores and 678 GB RAM memory. Persistent storage was provisioned as 3.5 TB of low latency storage and 30 TB of intermediate latency storage. An exome sequencing analysis pipeline was deployed on the swarm cluster as a stack of Docker services, namely FASTQ preprocessing (fastp), sequence alignment (bwa), BAM file processing (Sambamba), indel realignment (GATK), and variant calling (Varscan2). RESTful API endpoints were used for job submission and monitoring. NGS data from 5 well- characterized NIST reference materials and 5 pancreatic neuroendocrine tumors were analyzed using this setup. Average analysis time was 160 minutes/sample for 80x sequencing depth, using up to 92% and 55% of the CPU and memory resources, respectively. Analytic sensitivity for all variant types in the NIST reference samples was 94.6%.

Conclusions: Containers and orchestration technology can be used to provide a robust HPC environment for supporting genomics data analysis. In contrast to a traditional HPC setup, using containers can provide a more consistent and modular infrastructure to deploy and update data analysis pipeline components.

J Pathol Inform. 2019 Sep 16;10:28.

Patient Timeline: A Tool for Visualization of Objective Clinical Data

Jacob Spector 1, Erica Schleimer 1, Stephan Sanders 1, Riley Bove 1, Kate Rankin 1

Background: One of the many challenges of clinical medicine is trying to parse different sections of the electronic medical record to conceptualize a coherent clinical story for a patient. Information that clinicians find most relevant is frequently in separate areas of the medical record, and is often mixed with extraneous data. Not only does this increases the amount of time required to review a patient's chart, but it could also potentially lead to important results being missed. Thus we created Patient Timeline, a tool for efficient visualization of objective clinical patient data.

Methods: Patient Timeline is part of the UCSF BRIDGE project, which utilizes the Epic (Madison, Wisconsin, US) application programming interfaces to pull and visualize individual patient data. It is a web application built using a Flask backend and it uses Javascript and D3.js to create visualizations. The BRIDGE project and Patient Timeline are still in development, with plans to go live with the BRIDGE project in April of 2019.

Results: Patient Timeline [Figure 1] has multiple different views, each showing data at a different granularity. The “Preview” view shows the number of datapoints for each type of data on a log scale. The “Expanded” view shows different data types on the same page and has custom methods for showing each type of data. Laboratory panels, for example, are displayed as a rectangle and given a different color if all components are normal vs at least one component being abnormal. Clicking on this rectangle then creates a line graph showing the values of each component of the panel over time. We have worked closely with clinicians in several different specialties to determine what information they find most relevant. This has allowed us to create more meaningful visualizations that are specialty specific.

Figure 1.

Figure 1

Open in a new tab

Example of the Patient Timeline

Conclusions: Patient Timeline is a tool that brings together multiple types of data to create visualizations that allow clinicians to review objective clinical data about their patients. We believe that use of this tool could potentially save clinicans time, increase the quality of patient care, and alleviate some of the annoyances of using current electronic medical record systems.

J Pathol Inform. 2019 Sep 16;10:28.

Incorporating the Nonbinary Gender into the Laboratory Information System

Jennifer S Woo 1, Ryan O'Connell 1, Daniel Farrell 1, Nalini Merchant 1, Witoon Lee 1, Elizabeth Cookson 1, Vani Grosvenor 1, Tina Chuang 1, Carol Eade-Viele 1, Scott Keefer 1, Sherif Rezk 1

Background: In the United States, a growing number of states are legally recognizing the nonbinary gender. In compliance with new laws, hospitals are beginning to accommodate nonbinary individuals into the ADT. The incorporation of the nonbinary gender into the laboratory information system (LIS) is challenging as some tests have reference ranges established by biological male and female sex (sex-specific ranges), and most LIS have limited functionality in displaying more than one reference range. Abnormal flagging rules must also be defined for this patient population.

Technology: LIS: Sunquest v8 (Sunquest, Tucson, Arizona, USA), Electronic health record: Epic 2018 (Epic, Verona, Wisconsin, USA), Corepoint Integration Engine (Corepoint Health, Frisco, Texas, USA).

Methods: Since 1/1/2019, nonbinary patients are registered in our ADT as “unknown sex” (U). The U category is already built in our LIS, however no reference ranges were applied to any tests in this category. We sought to apply reference ranges to all tests for U patients. For tests with sex-specific ranges, we sought to display both male and female ranges, and flag results that extend beyond the male and female range overlap to favor higher sensitivity for recognizing true abnormal values.

Results: Labs with common male and female ranges were also applied to U. Gender-based calculations attached to result codes in the LIS were evaluated. Critical ranges were applied to all tests. We utilized an interface engine to apply male and female ranges to tests with sex- specific ranges and to apply flagging rules. Logic and tables were built in the interface engine to append the male and female ranges along with a comment to explain the rationale for the display of reference ranges and abnormal flagging. Logic was also used to evaluate the narrowest of the male and female ranges to provide abnormal flagging.

Conclusions: The incorporation of the nonbinary gender into the LIS poses a variety of laboratory informatics challenges. Our institution utilized an interface engine to display male and female reference ranges for tests with sex-specific ranges, and to provide abnormal flagging. Ongoing maintenance is needed to maintain the table as reference ranges change and tests with sex-specific ranges are added.

J Pathol Inform. 2019 Sep 16;10:28.

Application of Reverse Federated Database System for Clinical Laboratory Service

Keluo Yao 1, Christopher L Williams 2, Ulysses G J Balis 1, David S McClintock 1

Content: Traditionally, the electronic health record (EHR) system acts as a federated database system (FDS) that provides a uniform interface for end users to retrieve data from a multitude of separate dependent databases. Increasingly, the dependent databases and their originating clinical services require the full understanding of the patient's clinical history through the EHR in order to function. To fill the gap, numerous time consuming and error prone manual workarounds have been devised to allow the originating clinical services to function. Here we have created a novel web service application (WSA) for protein electrophoresis that applies reverse federation database system (RFDS) by interfacing the federated EHR as a dependent database to automatically gather information to streamline the laboratory workflow.

Technology: Web Application Programming Interface (API), Node.js v10, Representational state transfer (REST), Virtual Private Network (VPN), Docker (San Francisco).

Design: Using Node.js and Docker container, we created a server that can query EHR for essential laboratory values (e.g., total protein) through a Web API REST interface. The queries are initiated by the WSA based on the patient identity and information pulled from the protein electrophoresis instrument database. Additional clinical information including medications, documents, radiology, and other additional laboratory results are also queried and compiled. Security and access are achieved through VPN, Web API key, dedicated machine account, and end-user authentication. Additionally, the use of a server to handle EHR query reduces the exposure of EHR Web API to end user tampering.

Results: Figure 1 (next page) shows WSA UI pulling data from both EHR (A) and protein electrophoresis (B) and synthesizes a dashboard (C) and presents other pertinent information including medications (D), other laboratory results (E), and clinical documents (F). Preliminary alpha testing and feedback from the laboratory staff and pathologists indicates the WSA will significantly improve the workflow.

Figure 1.

Figure 1

Open in a new tab

The electronic health record (a), as a federated database system, allows multiple dependent databases to be accessed uniformly. The protein electrophoresis workflow (b) requires some information from the electronic health record and can treat it as a dependent database to provide a dashboard (c) medication reference (d), additional laboratory result (e), and clinical documentation (f)

Conclusion: We have successfully designed a WSA that applies RFDS to streamline the workflow of a laboratory service. The secure design and the opportunities it offers can be used on any clinical service that relies on low throughput and error-prone manual information extraction from the federated EHR for routine operation.

J Pathol Inform. 2019 Sep 16;10:28.

Efficient Breast Cancer Metastasis Detection from Histological Images Using Convolutional Encoder-Decoder Network

Xinyue Meng 1,2,*, Huan Liu 1,3,*, Hanbo Chen 1, Lei Guo 3, Hailong Yao 2, Xiao Han 1, Jianhua Yao 1

Background: Breast cancer is the most common cancer type among women. Whether the cancer metastasizes into lymph nodes is an essential factor for treatment planning. Cancer metastasis can be detected through examining H&E stained histological slides. However, the diagnosis is time consuming and small metastatic regions are prone to be missed by pathologists. We explore convolutional neural network (CNN) based models to efficiently detect breast cancer metastatic regions on both whole slide images and microscopic images.

Methods: In this work, we propose two variants of Linknet model (one kind of convolutional encoder-decoder neural network) for breast cancer metastasis detection. One is an improved ResNet-type Linknet model with skip connections (noted as LinkNet) and another is a smaller model (noted as LightLinkNet) which removes the uninformative feature map channels from the Linknet model to significantly reduce the amount of model parameters. Both models are compared with the classic UNet model [Figure 1]. We tested our methods on the publicly available dataset in Camelyon2016 (400 WSIs, 270 for training and validation, 130 for testing). We also transfer our models to work on live microscopic images.

Figure 1.

Figure 1

Open in a new tab

Detection on whole slide imaging (WSI) and microscopic images (a) WSI, (b) LinkNet output on WSI, (c) and (d) Light link net output superimposed on microscopic images

Results: At 10X magnification level, the detection performances in terms of area under Free Response Receiver Operating Curve (FROC) for LinkNet, LightLinkNet and UNet are 0.804, 0.794, and 0.748 respectively. The segmentation performances in terms of Intersection over Union (IoU) score for LinkNet, LightLinkNet and UNet are 0.886, 0.877, and 0.864 respectively. The running times on a 2048x2048 microscopic image for LinkNet, LightLinkNet and UNet are 0.165, 0.071 and 0.231 seconds respectively. The model sizes for LinkNet, LightLinkNet and UNet are 87.3, 4.8 and 31.4 MB respectively.

Conclusion: The improved LinkNet model achieves the best performance in detecting and segmenting breast cancer metastatic regions in lymph nodes. The proposed LightLinkNet model achieves real time detection rate under 10x microscopic examination. Our model can be applied in clinical setting to assist pathologists detecting metastasis efficiently.

J Pathol Inform. 2019 Sep 16;10:28.

Effective Text Searching in Pathology Records: Powerful Solutions That are Available to Every Pathologist

Yonah C Ziemba 1, Feifan Chen 1, Kingsley Ebare 2, Tarush Kothari 1

Background: Difficulty in “datamining” is a barrier encountered in many projects that involve finding cases based on text of surgical pathology reports. Sophisticated language processing tools are usually not available, and simple search for free text or ICD codes are not accurate enough to fill the needs of many projects. Here, a method is presented that uses only Microsoft Excel formulas.

Methods: As illustrated by the examples in Table 1, the approach is based on five strategies: (1) Edge testing recognizes that in order to craft an accurate set of search terms, it is necessary to know all the language variations used in the pathology records. Therefore, it is helpful to begin with 2 searches and focus specifically on the discrepancies. (2) Proximity searching considers context, such as “residual radial scar” in proximity to “is not seen”. (3) Sensitivity/specificity testing involves a small, manually reviewed sample to evaluate whether a given set of terms is too broad or too narrow. (4) Training sets/validation sets recognizes that search terms designed on a training sample are likely to be “overfitted” to the training sample and have less accuracy in the broader dataset. Therefore, validation is accomplished by measuring accuracy on a new sample that was not used during training. (5) Exclusionary rules are developed by incorporating exclusions into the search terms.

Table 1.

Examples of the five strategies

Edge testing
Origin of concept: Used in training students of technical fields, where attention is focused on challenging variants by identifying cases that yielded discordant results in a previous set of students. In the context of pathology reports, it is helpful to begin with a term that is overly broad and another term that is overly narrow and review the discrepancies to discover language variants. For example, a review of cases that match “radial” but do not match “radial scar” will reveal variants such as “radial sclerosing lesion”
Technical instructions: The formula=ISNUMBER (SEARCH(“radial”, A2)) is entered into column B, = ISNUMBER (SEARCH(“radial scar”, A2)) is entered into column C, and=B2=C2 is entered into column D. Special attention is focused on cases were column D returns “False”
Proximity searching
Origin of concept: A free text search strategy used by lawyers to identify content of interest in the review of thousands of documents. When a search term may occur thousands of time in an irrelevant context, proximity searching will identify the occurrences that are specifically in the context of interest.
Technical Instructions: The formula = (ISNUMBER (SEARCH(“is not seen”, MID (A2, SEARCH(“residual radial scar”, A2),50)))) will return True only in cases where “is not seen” occurs within 50 characters after “residual radial scar”
Sensitivity and specificity testing
Origin of concept: Used to evaluate medical diagnostic tests
Technical instructions: A random set of 100 records is evaluated manually as “true” or “false” with respect to the inclusion criteria. The formula is applied to an adjacent column. When the formula returns “true” on a case that was manually evaluated as “true”, the occurrence is designated a TP. When the formula returns “true” on a case that was manually evaluated as “false”, the occurrence is designated a FP. A similar process is used to designate TN or FN. Calculate sensitivity=TP/(TP+FN) and specificity=TN/(TN+FP)
Training set and validation set
Origin of concept: Used to construct algorithms in machine learning
Technical instructions: Use formula=RAND() to assign a random value to each record, and sort by this value. The top set of 100 values can serve as a randomized training set, and the next set of 100 values can serve as a randomized validation set
Exclusionary rules
Origin of concept: Since criteria for research projects include both inclusions and exclusions, it is timesaving and efficient to incorporate the exclusions directly into the search terms. For example, a search for cases that contain “radial scar” but do not contain “carcinoma” will avoid cases that need to be excluded based on project criteria
Technical instructions: The formula=AND((ISNUMBER (SEARCH(“radial scar”, A2))), NOT((ISNUMBER (SEARCH(“carcinoma”, A2))))) will return True only for cases that contain “radial scar” but do not contain “carcinoma”
Open in a new tab

TP: True positive, FP: False positive, TN: True negative, FN: False negative

Results: To test this approach, we compared three search methods for a project that required cases of “pure” radial scar found on core needle biopsies. Simple free text search for the term “radial scar” returned cases of mastectomies and co-existing high risk lesions and did not return cases of “radial sclerosing lesion.” Search by ICD codes returned all benign lesions. Accuracy of both of these methods was below 60%. However, using the strategies above, a set of rules based on free text search was developed that had greater than 96% accuracy.

Conclusion: This method is more accurate and effective than other options available to most pathology projects. All these strategies can be adapted in many applications, such as Excel, and provide great utility for many needs of a pathology department.

J Pathol Inform. 2019 Sep 16;10:28.

An Artificial Intelligence Approach to Variant Calling of Anaplastic Lymphoma Kinase Resistance Mutations in Clinical NGS Data

Jochen K Lennerz 1, Emily Chin 2, Marguerite Rooney 2, Mick Zomnir 1, Lev Lipkin 1, A John Iafrate 1, Long P Le 1, Alice Shaw 2

Background: ALK tyrosine kinase inhibitors (TKIs) are effective in treating advanced anaplastic lymphoma kinase (ALK) fusion-positive non-small-cell lung cancers (NSCLC) and specific ALK variants are associated with the development of resistance to specific TKIs. Humans struggle to harness the full potential of the highly complex next-generation sequencing bioinformatics pipeline output. As a consequence, the decision to report a variant remains difficult and we considered the discrete nature of the data and the binary decision (report vs. not-report) as an ideal setting to apply an artificial intelligence (AI) approach for variant reporting.

Technology: The random forest model was derived from independent datasets (training and validation) capturing the reporting decision on >36,000 variants with ~500 features per variant resulting in a matrix of >18 million data points. The model output is a contiguous prediction score from 0 (not report) to 1 (report) and a visual drill-down functionality allows exploration of the underlying features that contributed to the decision.

Methods: We assessed diagnostic performance of an AI model in calling ALK-resistance mutations in n=50 consecutive ALK fusion positive patients who relapsed on TKI-therapy and underwent repeat biopsy at MGH.

Results: We examined ML-data of n=76 tests from n=50 patients with a total of n=130 reported variants (and =115 not reported variants) included a total of n=31 ALK point mutations. Setting a screening threshold of the model at >10% for reporting showed only one false-negative (p.Ile171Asn) variant and 96.7% sensitivity. The average model score for ALK variants was 0.664 (range: 0.08–0.98; median 0.8) and did not show significant differences from other reported variants (0.636; 0–1; 0.7; t-test 0.66). The model would have called n=18 of the non-reported control variants (average 0.07; range <0.001–0.64; P<0.0001) and was 84% specific. Review of the drill-down function identified prior call frequency, allelic ratio, and predicted transcript consequences as common model features. Importantly, the model is currently agnostic to the medical literature and does not take clinical parameters (e.g. TKI type) into account, which may further improve performance.

Conclusion: We consider the application of artificial intelligence to discrete NGS datasets as one approach to identify relevant variants in the setting of NGS genotyping.

J Pathol Inform. 2019 Sep 16;10:28.

Visual Risk Pattern Recognition Informatics Expedites Rapid Triage of High Risk Complex Patients Vulnerable to Diagnostic Adverse Events

Eleanor M Travers 1

Background: The rising rate of adverse diagnostic events (ADxE) is a clinical challenge that needs Laboratory Medicine's innovative, proactive reporting of patient risk pattern severity staging for rapid clinical action. They are major causes of poor outcomes and death in complex patients with multiple critical organ/system dysfunction. Patients with multiple diseases and polypharmacy produce patterns of multiple, significantly abnormal biomarkers. The earliest scientific diagnostic “sentinel indicators” of dysfunction causing diagnostic adverse events are patterns of significantly abnormal, severity - graded biomarkers.

Technology: A new, novel use for Laboratory Medicine informatics innovation is Figure 1, The New Patient – Specific Report.

Figure 1.

Figure 1

Open in a new tab

The new patient specific report

Methods: Signal Detection Theory: Signal detection theory is used to find optimal combination of diagnostic test results is a means to quantify critical information-bearing patterns vs. “noise”

Predictive Diagnostic Analytics: Recognizes high risk patterns in diagnostic data used to accelerate clinical physician's decision- making for rapid triage of risk patients most vulnerable to ADxE.

Graphical Representation of Numerical Data: Laboratory Medicine biomarker data visualization converts numeric biomarker data into patterns of clinical knowledge. Visual informatics report formats with diagnostic “trigger signal” patterns deliver patient - specific, severity -graded critical data by translating numbers into visual images using pattern recognition.

Results: Conclusions: Laboratory Medicine services report digitized biomarker information without interpreted visual graphics that assist physicians realize the clinical value of these data. Early Decision Support (EDS) systems using visual graphics images of patient's risk severity patterns of significantly abnormal biomarker test results are needed “just – in- time to take action to prevent an adverse diagnostic event.

J Pathol Inform. 2019 Sep 16;10:28.

Using Generative Adversarial Networks to Remove Unwanted Pen Marks from Scanned Histology Slides Images

Fangyao Hu 1, Cleopatra Kozlowski 1

Background: In order to perform digital image analysis on scanned pathology slides, the algorithm has to be designed to either ignore, or digitally correct for various artifacts on the image, that are irrelevant for biological interpretation. One common artifact is the result of manually drawn pen marks on physical histology slides, often employed by pathologists to identify relevant regions of the slide (for example, tumor lesions). Often, the tissue under the marked area contains important information that we do not wish to discard. Thus we devised a method to train an algorithm to almost perfectly remove markers from Hematoxylin and Eosin stained lung tissue slides.

Technology: For this purpose, we employed generative adversarial networks (GANs), which consist of two neural Networks that 'compete' (hence the term, adversarial) with each other to maximize the accuracy where a new image is 'generated' from an original.

Methods: We obtained a dataset of 105 unmarked hematoxylin and eosin-stained lung tissue slides, and scanned them at 20x. We then randomly marked these slides with red, blue, black, and green markers and rescanned the slides. We aligned the original and marked images with SlideMatch software (Microdimensions, Germany). We produced 500x500 pixel tiles areas of regions were marking were present using Definiens Developer software (Definiens, Germany). We then selected a subset of these slide pairs (the 'training' set), and used the GAN approach to train a network that generates the clean image from the marked image. The trained GAN was applied to only the marked images in the 'test' set, to see how much they differed from the unmarked test images.

Results: Inspection of the test dataset showed a very high correspondence between the GAN-generated images on the marked images, and the original, unmarked images [Figure 1].

Figure 1.

Figure 1

Open in a new tab

The pen-marked image (left), the generative adversarial networks -generated image (center), and the original unmarked image (right). The difference between the generated image and original is almost imperceptible

Conclusions: Our proof of concept work of using GANs for marker removal showed remarkably accurate results. This method is highly promising in uses of other artifact correction tasks in whole slide imaging.

J Pathol Inform. 2019 Sep 16;10:28.

A Modern Approach to Specimen Tracking in a Geographically Distributed Department

Joshua Jacques 1, Brian Royer 1, John Hamilton 1, William Hubbard 1, Todd Kandow 1, Amy Mapili 1, David McClintock 1, Ulysses Balis 1

Background: Specimen tracking in the current generation of Laboratory Information Systems (LIS), while useful at a single laboratory location, many times do not have functionality allowing for tracking across separate facilities. Specifically, when specimens have to be grouped (binned) and shipped at varying temperatures to multiple locations, vendor-based tracking systems are lacking. This gap in functionality became readily apparent when our clinical laboratories were split 3 miles apart during and following a major renovation and relocation project – a review of current vendor systems made it clear that, in order to achieve the desired tracking functionality, we would need to build our own system (PathTrack).

Technology: Node.js v10, Webpack v4, React.js v16, Redux.js v4, Ant Design v3, Oracle v.12, Sockets.

Design: The basic design goals for PathTrack were: 1) easily accessible, 2) versatile, 3) data-driven and 4) very responsive for users. It was built as a web-based application using React, driven by an Oracle database housing all business logic. All computational heavy lifting is performed server side with the user interface side quickly rendered by React, which is fed data using socket technology. Additionally, data from the LIS was integrated through Oracle, allowed specimen specific information to be accessed real-time and for existing LIS barcodes to be used. Through this architecture, we eliminated the need for a bidirectional interface with our LIS as all tracking data is updated and stored within PathTrack – double barcode scanning is NOT required to receive specimens in both applications.

Results: From Go-live in June 2018 to March 2019, over 500,000 specimens have been tracked in PathTrack. Screenshots of the application are shown in Figure 1. Importantly, the lost specimen rate has not changed significantly after splitting the workflow between two separate sites [Figure 2].

Figure 1.

Figure 1

Open in a new tab

Path track interface

Figure 2.

Figure 2

Open in a new tab

Lost specimens before and after department separation

Conclusions: Designing your own specimen tracking system is a large undertaking, however it also has significant benefits. With only minor incremental work by lab staff we have been able to efficiently keep track of specimens between two separate sites, including maintaining proper temperature regulation. Further, labs can now: 1) view specimens in route, 2) batch testing based on incoming specimen volumes, and 3) receive alerts when irregular workflows are attempted (i.e. trying to route a specimen to the wrong lab and/or to the wrong site). Additional work and workflow modifications are ongoing, with expansion of the system now moving into the ambulatory clinics in 2019 and off-site locations in 2020.

J Pathol Inform. 2019 Sep 16;10:28.

A Collaborative Project to Produce Regulatory-Grade Pathologist Annotations to Validate Viewers and Algorithms

Brandon D Gallas 1, Mohamed Amgad 2, Weijie Chen 1, Lee A D Cooper 2, Sarah Dudgeon 1, Hannah Gilmore 3, Anant Madabhushi 4, Roberto Salgado 5, Joel Saltz 6, Ashish Sharma 2, Darren Treanor 7,8,9, Jithesh Veetil 10, Bethany Williams 7, Kyle J Myers 1

Background: Given one digital pathology system has made it on the market, other companies are expected to follow. Next to come are image analysis and artificial intelligence algorithms. Research on these algorithms is ubiquitous, and challenges are being organized to share images and annotations to stimulate development and understanding.

Methods: The FDA is leading a team to collect regulatory grade data (images and annotations). The use case that we chose to pursue will identify, quantitate, and spatially characterize breast cancer tumor infiltrating lymphocytes (TILs) in tumor and stromal regions. The primary annotations will be collected using the microscope: the reference technology that is not limited by any digitization of color or space (in and across focal planes). We have a dedicated system that uses a microscope, motorized stage, and camera that allows us to register the locations of the annotations so that they may be mapped onto any digital scan of the glass slide. The system was developed at FDA and is named eeDAP, evaluation environment for digital and analog pathology (https://github.com/DIDSR/eeDAP). We will collect data at pathology conferences and other settings where we can efficiently employ a high-volume of pathologists. We will also collect digital-mode data through a web-based viewer to allow comparisons between microscope- and digital-mode annotations. We will develop statistical methods to assess/assure quality of the annotations. We will pursue an FDA Medical Device Development Tool (MDDT) qualification for a substantial subset of collected data, with some data to be made public. MDDT qualification allows algorithm developers to have confidence that the data can be used in medical device submissions to the FDA. Using the data, we will also pursue an MDDT or an open or mock 510(k) submission for a WSI viewer and for an algorithm. The public data can be used to facilitate algorithm development ahead of validation.

Results: We are currently creating the project structure: workgroups and leadership. We will provide a project update at the conference.

Conclusions: The project is open to the community at https://nciphub.org/groups/eedapstudies/wiki/HighthroughputTruthingYear2 so that the community can participate or follow the regulatory paths we will demonstrate.

J Pathol Inform. 2019 Sep 16;10:28.

Revisiting Whole-Slide Imaging in the Context of Big Data: Strategies for Data Archival and Retention

Mark D Zarella 1, Jennifer Jakubowski 1

Background: Whole-slide imaging (WSI) not only supports clinical workflows but also serves as a substrate for research, teaching, and the development of deep learning algorithms for computer- assisted diagnostics. Despite the clear utility in compiling large data sets for these purposes, many institutions are faced with significant questions about image retention because of the considerable amount of data storage required. We review methods in WSI storage and curation, with particular emphasis on image compression and the unique tradeoffs that exist in WSI between file size and bandwidth.

Technology: We examined WSIs acquired at our institution over the course of 10 years and across multiple slide scanners.

Methods: We modeled the impact of conventional image compression and pyramidal representation, and developed an optimization algorithm based on real-world WSI viewing. In addition, we present a novel method of file compression for archival purposes using inter-slide compression.

Results: Guided by natural patterns of whole-slide viewing measured in subjects as they navigated virtual slides, we established a tunable strategy for pyramidal representation of whole-slide images that optimizes the tradeoff between bandwidth utilization and file size. Notably, for strictly archival purposes, this strategy can easily be augmented to reduce whole-slide file sizes by an additional 10-20% without a corresponding loss of information. We considered an alternative framework strictly for image archival by exploiting the natural redundancy across WSIs and the intrinsic properties of histology images. Of the images tested, we found that serial sections could be compressed using this technique with a negligible loss in image information. We demonstrated that this technique performs exceptionally well on cytology WSIs collected at multiple focal planes (z-stacks), achieving a reduction in file size of a factor of 4 or more without sacrificing image quality even at 40x magnification. Together, these techniques result in an overall reduction in WSI file size of 20-80% without a corresponding loss in image quality.

Conclusions: To support an indefinite retention policy, archival requires a different image storage and compression strategy than conventional WSI approaches. Substantial reductions in overall file sizes are possible by employing the strategies we propose, with particular improvements observed in z-stacks and serial sections.

J Pathol Inform. 2019 Sep 16;10:28.

Atlas of Digital Pathology Database: Pathologist Supervisory on Recognition of Histological Tissue Types

Mahdi S Hosseini 1,2, Lyndon Chan 1, Corwyn Rowsell 3,4, Konstantinos Plataniotis 1, Savvas Damaskinos 3

Content: The field of computational pathology is slowly adopting recognition tasks from computer vision due to lack of well-labelled databases for supervised training. Existing datasets provide limited diagnostic information at the slide-level with unstructured textual form of tissue labels and diseases. We have sought to address these shortcomings by compiling a new prototype of Atlas of Digital Pathology (ADP) database which is labeled for diverse Histological Tissue Type (HTT) in image patch-level.

Technology: A total of 100 anonymized glass slides (mainly H&E) were selected (each sized 1”x3” and 1.0 mm in thickness) and digitized using Huron TissueScope LE1.2 Whole Slide Image (WSI) scanner at 40X magnification (0.25um/pixel resolution, uncompressed TIFF files). Each WSI was then divided into a randomized subset of recognizable non-background patches of size 1088x1088 pixels. In total, 17,688 patches were collected.

Methods: Multi-label class of tissues are defined for patch labeling. A generalized hierarchical taxonomy of 57 HTTs are defined to cover broad range of histological tissues. A total of five labelers were trained based on this taxonomy to perform labeling. A random set of 1,000 tissue patches was reviewed by a board- certified pathologist where an excellent concordance between original labelers and the Pathologist was reported. The quality of ADP labels is further evaluated by training three different Convolution Neural Networks (CNN) for classification. A visual attention aid tools is developed to assist pathologist for searching certain tissues at WSI level.

Results: Table 1 shows the test results using three CNNs trained on ADP. The following Figure 1 shows the test results using three CNNs trained on ADP. The demography of all HTTs within similar patches are shown in Figure 1a for the ADP database. Few ROC curves of recognition performance using VGG16 network are shown in Figure 1b. An example of heatmap representation of the Exocrine Glands (G.O) on WSI patch-level are shown in Figure 1c.

Table 1.

Test results using three CNNs trained on ADP

CNN classifier F1-score
VGG16 Net 0.8968
ResNet18 0.8657
Inception-V3 0.8950
Open in a new tab

CNN: Convolution neural network

Figure 1.

Figure 1

Open in a new tab

(a) Co-occurance of histological tissue types for atlas of digital pathology; (b) receiver operating characteristic on five histological tissue types; and (c) visual attention aid heatmap

Conclusions: Identification of different HTTs is feasible by compiling an annotated database of WSI image patches across different organs and training a supervised machine learning tool for tissue recognition. Preliminary results suggest that ADP can be highly useful for development of automated tools of computational pathology to assist pathologist in clinical applications.

J Pathol Inform. 2019 Sep 16;10:28.

Quantifying Data Element Evolution in Three CAP Electronic Cancer Checklist Templates

Keren I Hulkower 1, Eric M Daley 1, Richard Moldwin 1

Background: We studied the guideline changes that occurred during the maintenance of the College of American Pathologists (CAP) electronic CAP Cancer Checklists by looking at eCC releases and data element changes over the past 10 years. This period reflects the transition from American Joint Committee on Cancer (AJCC) Staging Manual from the 7th to 8th editions, as well as the introduction of the enhanced eCC and SDC XML formats.

Technology: The eCC XML Comparator is a web-based tool for identifying differences between eCC XML template versions. SQL queries were executed against the primary eCC development database to review change comments made by eCC template modelers. The Project Tracker database is used to track change documentation for eCCs.

Methods: Using the above tools, change tables were generated for each eCC release to compare all data element and metadata changes between versions of each template. We reviewed all version-specific changes to the Breast, Colon and Lung eCC templates that were made between 2009-2019. All changes were compiled into a database, partitioned by change type, and organized by eCC template versions.

Results: XML Comparator change data for Colon alone reveal 899 total changes, 327 added items, 261 deprecated (removed) items, and 311 minor changes over the 14 releases of this template. In addition, the eCC XML format changed from “legacy” to enhanced” to “SDC” format during this period. Modeler-entered change comments (total 426) were obtained by SQL query against the eCC database. This query produced 125 comments on added items, 79 comments on deprecated items, and 222 comments for all other changes. Of the 426 modeler-entered change records, 423 had Project Tracker records documenting the change deliberations.

Conclusions: The development of practice guidelines for cancer reporting is an expensive and time-consuming process with significant clinical, quality, regulatory, research and financial impact. The magnitude of changes identified in this report underscores the need for centralized management and long-term support to maintain content, informatics infrastructure, and rollout assistance for this type of rapidly changing practice guideline.

J Pathol Inform. 2019 Sep 16;10:28.

Application of Artificial Intelligence for Automatic Evaluation of Routine Whole Slide H&E Images for Presence of Malignancy

Anthony Magliocco 1, Eliron Amir 2, Lotan Chorev 2

Background: Convolutional Neural Networks have been applied for various histological interpretation tasks in past years. Most academic work that presents the application of such technology is limited to a specific indication only. Here, we validate the use of a novel proprietary General Purpose Histology Artificial Intelligence Framework (GPHAI) based on Convolutional Neural Networks in detecting malignancy and neoplasm in multiple common indications.

Methods: We evaluated the performance of a GPHAI (Nucleai-AI-Framework, v1.03) developed by Nucleai Ltd. The framework was trained to distinguish between benign and malignant/dysplastic cases using whole slide images (WSI) of routine H&E histology slides from 3 indications: 2512 Colonic polyps, 2038 Breast core biopsies and 3007 Prostate core biopsies. We performed a retrospective analysis of a test sample sets taken from a laboratory clinical archive containing 200 colon polyps WSI, 200 breast core biopsies WSI and 200 prostate core biopsies WSI. All cases were evaluated by multiple (n=2) pathologists. The sample sets contained equal distribution of benign and malignant/dysplastic. Colonic polyps slides were selected from the following categories: normal colonic mucosa, hyperplastic polyps, adenoma with low grade dysplasia, adenoma with high grade dysplasia. Prostate and Breast slides that were included in the study were selected such that IHC staining was not required to produce a binary (benign / malignant) diagnosis.

Results: The overall concordance between the GPHAI and the reference pathologists was 98.5% for Colon Polyps (sensitivity 98.8 and specificity 98.2), 97% for Breast core biopsies (sensitivity 98.03 and specificity 95.9) and 97% for Prostate core biopsies (sensitivity 97.01 and specificity 96.7).

Conclusion: The findings demonstrate the capability of the investigated GPHAI to successfully detect malignancy in various tissue types and indications. This shows the potential of a novel artificial intelligence system in assisting in clinical diagnostic workflows including triaging cases or quality control. Sources: Nucleai Ltd, Tel Aviv, Israel.

J Pathol Inform. 2019 Sep 16;10:28.

Identification of Basal Cell Carcinoma in Intro-operative Frozen Section Using Deep Learning on Smart Phone Images

Junyan Wu 1, Eric Z Chen 2, Jingwei Zhang 3, Jay J Ye 4, Limin Yu 5

Background: Basal cell carcinoma (BCC) is the most prevalent cutaneous malignant neoplasm. Its incidence rises at a rate of 3–10% annually. Surgical excision is a common, highly effective treatment for the management of BCC. Mohs surgery is a specialized surgical technique that minimizes the amount of normal tissue resected and simultaneously optimizes control of the tumor margins, is commonly performed when tissue sparing is desired due to cosmetic or functional concerns. Accurate intra-operative frozen section (FS) is critical during Mohs surgery, which provides rapid gross and microscopic analysis of a specimen. FS of BCC specimens can present a significant challenge to pathologist dependent upon his/her experience and training. Frequently, consultation of a dermatopathologist is requested by the FS pathologist. Patient care might be compromised when availability of dermatopathologist is limited.

Technology: Machine learning and artificial intelligence techniques have made breakthroughs in health care. Convolutional neural network (CNN) is a class of artificial neural networks that is efficacious in various computer vision tasks, including pathology and radiology images.

Methods: We examined the use of CNN in identifying BCC during FS in images of smart phone microscopic photography. Smart phones were used to take histologic FS images of BCC and normal tissue including adnexal structures at 20x and 40x magnifications. The subtypes of BCC studied includes superficial, nodular, micronodular and infiltrative types. The smartphone images were labeled by three dermatopathologists, if BCC present, and randomly assigned to training (478 positive and 514 negative images) and testing (120 positive and 129 negative images) groups. The images were further sliced to non-overlapping 256 x 256 pixels patches. Patches from positive images containing tumor tissue > 95% were used as positive patches; otherwise, they were discarded. Patches extracted from negative images were defined as negative patches. Consequently, the training dataset contained 30708 positive patches and 30708 negative patches; the testing dataset contained 7536 positive patches and 7535 negative patches. Resnet50 and Densenet161, pre-trained on ImageNet, were used as deep learning platforms for BCC identification. Class Activation Maps (CAM) was also implemented for visualization of heat map.

Results: Performance of accuracy=0.90, precision=0.9, recall=0.9 was achieved using Resnet50 when positive patch cutoff is greater than 95%. Performance of accuracy=0.905, precision=0.92, and recall=0.90 was achieved using Densenet161 when positive patch cutoff>95%.

Conclusion: CNN is a promising technology that can help pathologists improve diagnosis accuracy during BCC frozen section, particularly when expert consultation is needed.

J Pathol Inform. 2019 Sep 16;10:28.

Prostate Cancer Diagnosis and Quantification Using AI-enabled Software (SW)

Wei Huang 1, Samuel Hubbard 1, Parag Jain 2, Ramandeep Randhawa 2

Background: Morphology is the gold standard for prostate cancer diagnosis. Evaluation of prostate biopsy slides is a time-consuming process. High inter- over variability in Gleason scoring has been reported: 40% discordance between general and sub-specialty pathologists. Digital pathology has gradually gained foothold in pathology practice. Yet the problems associated with manual scoring and quantification remain. A universal and standardized platform for Gleason grading and Gleason pattern (GP) quantification trained by GU pathologists is needed to achieve accurate and reproducible diagnosis.

Materials and Methods: One thousand prostate biopsy cases, including all Gleason Grade Group (GGG) cancers were selected from the pathology archive at the University of Wisconsin-Madison, and were scanned with Aperio CS2 (Leica) at 40x to create digital slides. The slides were split into a training set and a test set. The training slides were then annotated by the GU pathologist (WH) to assemble a balanced dataset of varied morphologies, including GP3, GP4, GP5 cancer, high-grade prostatic intraepithelial neoplasia (HGPIN), perineural invasion (PNI), vessels and lymphocytes. The team used this data to train their deep learning architecture, which comprises multiple Deep Convolutional Neural Networks that are a combination of classification and segmentation networks. This architecture was fine-tuned to be sensitive to very small amounts of high grade cancer. The trained software auto-annotates the entire WSI into the various cancer and benign pattern groups, and provides summary statistics of Gleason score, quantification of cancer area, and the percentage of each cancer pattern. The software was validated on a test set of 200 biopsy slides spread over the various cancer grade groups to establish concordance with the GU pathologist (WH).

Results: At this stage, the AI-enabled Software showed high efficiency and precision in cancer scoring and quantification: It takes one click, less than 3 minute to score and quantify cancer using the software, even with minor reassigning Gleason pattern for a few acini by the user [Figure 1]. The overall accuracy in Gleason scoring cancer is 95% [Figure 2].

Figure 1.

Figure 1

Open in a new tab

Example of AI-enabled software grading of Gleason scoring for prostate cancer

Figure 2.

Figure 2

Open in a new tab

Comparison of pathologist vs. software grading for Gleason scoring of prostate cancer

Conclusions: Deep learning enabled cancer-grading software offers objectivity, greater efficiency and precision in prostate cancer scoring and quantification. It has potential to help pathologists to minimize inter-observer variability and to increase efficiency in their practice.

J Pathol Inform. 2019 Sep 16;10:28.

Analysis of Cell Galleries as an Interface for Reviewing Urine Cytology Cases

Adit B Sanghvi 1, Erastus Z Allen 1, Robin L Dietz 2, Keith M Callenberg 1, Liron Pantanowitz 2

Background: Cytology, unlike histopathology, poses several unique challenges for deep learning algorithms ranging from image acquisition to data output. Digital cytology slides can contain over 100,000 variable 3D cells that need to be analyzed. A major challenge when applying a deep learning algorithm is how to best present the output to cytology end-users. Our aim was to develop a deep learning algorithm for urine cytology and use an image gallery to optimize the display of the results to cytologists.

Technology: 757 glass slides of urine samples were prepared using ThinPrep (Hologic, MA) and scanned by an Aperio AT2 scanner (Leica, CA) at 40x magnification. Resulting whole slide images (WSI) were de-identified and analyzed in a HIPAA-compliant cloud computing environment. Deep learning methods for classifying, analyzing and summarizing the cells for each slide were built in Python.

Methods: We built a deep learning algorithm that analyzed urothelial cells according to the Paris System for Reporting Urinary Cytology. Cell galleries were constructed by selecting the most suspicious cells according to our algorithm, randomly rotating and flipping images, and shuffling the display order to allow for WSI reuse. We simultaneously varied the number of cells and cell ranking algorithm in order to test both of these parameters. Cell galleries were reviewed and annotated by two pathologists via a web-based interface.

Results: A total of 1178 cell galleries were constructed from 757 WSIs [Figure 1]. Median gallery review time was 6.3 seconds (standard deviation of 5.8). Accuracies of 90.3%, 91.8%, and 85.4% were found for 12-, 24-, and 48-cell galleries, respectively, that were achieved on correlation with cytopathology diagnosis.

Figure 1.

Figure 1

Open in a new tab

Gallery of 24 urothelial cells with the “most atypical” PARIS features from an HGUC case

Conclusions: Galleries of the most relevant cells are a valuable format for review of urine cytology cases. Larger gallery size was correlated with an increase in the atypical call rate, while a moderate gallery size of 24 cells correlated most with the original cytopathology diagnosis. We recommend using image galleries that display 24 individual cells for an optimal and quicker review of cytology image analysis results.

J Pathol Inform. 2019 Sep 16;10:28.

Effective Use of R Language in Anatomic Pathology - Showcasing Diverse Examples of Usages Outside Statistical Computing

Jay Ye 1

Background: R is a free software environment not only with strength in statistical computing by also with capabilities of a general-purpose programming language. The documented usage of R in anatomic pathology has thus far been limited.

Methods: Various methods of R have been used, including natural language processing, interactive web application building, literate programming, and deep leaning. R programs directly interact with the data in the pathology database.

Results: (I) Interactive web applications built with R Shiny package have been routinely used by pathologist assistants to reduce anatomic pathology reporting errors and to assist coding staff in CPT coding. Web applications enable users to interact with data without knowing how to program in R. (II) Literate programming with R Markdown package has been used to obtain/summarize information from various aspects of pathology data, such as texts (melanoma depth, number of lymph nodes submitted in colectomy specimens, characters of early stage lung cancer, plotting geographic distribution of cancer case, etc.), utilization of special studies, case progression through the system, QA/PI monitors, and so on. The products of literate programming can be Word documents, PDF documents, interactive HTML documents or PowerPoint presentations, all consisting of interweaving narratives and computing results by R.

Conclusions: R computing environment is capable, versatile and accessible to the non-computer-science professionals. Awareness of these strengths of R and seeing some examples of its effective usages will likely promote the use of R by the anatomic pathology community.

J Pathol Inform. 2019 Sep 16;10:28.

Specimen Tracking from Operating Rooms to Pathology Intake

Peter Gershkovich 1

Background: Despite much effort to deliver diagnostic tissue from surgery to pathology labs, mistakes keep happening. Specimens are still getting lost or mishandled. The effect of these errors on patient care, although rare, can be detrimental. The methods for intradepartmental tracking of assets in Pathology can be used to reduce tracking errors outside Pathology.

Technology: Pathology Orders in Transit system (POinT) was developed in-house using the Java programming language. It is deployed as a web app on an OS X server with Apache Tomcat servlet container. The app communicates with the Epic EMR via HL7 message exchange enabled by the HAPI API based Hermes system. A MySql database was used for the persistence layer.

Methods: The Epic's surgical pathology order interface has been modified to trigger an ORM HL7 message every time specimen labels are printed. Each generated message, upon it's receipt in Pathology, is processed and saved to the database. The POinT app was equipped with a user interface for scanning specimen container barcodes. After each scan the POinT matches containers to the received orders and sends a physical delivery confirmation messages back to Epic. Recorded events are displayed on the color-coded dashboard showing specimen's lag and status.

Results: All resected tissue for Pathology now requires users to create an electronic order and print individual labels for each container. Each label contains a barcode with unique ID that enables tracking. A confirmation of the receipt of diagnostic tissue is now available in the EMR. Operating Room staff is able to monitor the receipt of each individual container and rapidly detect mishandling. The system enables users in the lab to automatically populate the LIS with patient information, collection dates, and the description of each part received.

Conclusion: The POinT helps to prevent a subset of specimen-handling errors by tracking the specimens from the moment of biopsy or resection to their arrival in the lab. It saves time and improves accuracy in Pathology by enabling specimen monitoring and automated data transfer during specimen accessioning. The implementation process revealed complex tracking issues. More work is required for creating an entirely paperless workflow.

J Pathol Inform. 2019 Sep 16;10:28.

An Innovative Web Application for Optimizing Pathologist Workflow in Clinical Pathology Sign-Out

Keluo Yao 1, Christopher L Williams 2, Ulysses G J Balis 1, David S McClintock 1

Background: Unlike in Anatomic Pathology, Clinical Pathology systems have not focused on the pathologist's needs when signing out. Protein electrophoresis is an example as the workflow relies on mounds of paper and hours of manual data entry. However, by leveraging recent advances, we have created a web application that can display all of the pertinent information for protein electrophoresis sign-out within a web browser. The user interface (UI) is novel in that it breaks away from the “single page application” paradigm and instead deploys a “drawer” system with multi-window support to account for multiple situations, such as the in-lab limited single monitor workstation environment as well as multi/large monitor workstations.

Technology: D3.js v4, Webpack v4, React.js v16, Redux.js v4, Ant Design v3, scalable vector graphics (SVG), Document Object Model (DOM).

Design: The UI is built using React.js to allow all UI elements to be componentized and reusable. Redux is use to track and update the component states and data. All graphics (e.g., waveforms) are created by D3 and passed to React for DOM manipulation. We used Ant Design to provide a grid and prefabricated UI elements such as buttons, forms, and drawer. All data are rendered in individual drawer pages that can be opened, closed, or rendered in a separate window using React Portals.

Results: Figure 1 shows the possible UI end user configurations, including on single (a and b), dual (c), and large monitors (d). All data are displayed on cascading drawer pages or as separate windows with seamless coordination. Alpha testing shows the application is able to render in real time without any noticeable delay.

Figure 1.

Figure 1

Open in a new tab

The protein electrophoresis web dashboard application is designed accommodate a wide array of screen monitor setups. When the work station is limited to a single standard sized screen, shuffling different drawer pages (a and b) will bring up the needed information. With two monitor screens, two pages can be viewed concurrently (c). The number of concurrent pages can be increased further on larger monitor screens (d)

Conclusion: We have successfully designed a web-based UI that allows single or multi-window operations to optimize sign-out workflow for protein electrophoresis. While currently limited to a single clinical pathology application, this technology can be retooled easily to support multiple clinical laboratory tests to allow adaptive use of monitor screen real estate and maximization of pathologist/laboratory personal productivity.

J Pathol Inform. 2019 Sep 16;10:28.

Adaptive Automated Gross Transcription Using Phrase Express: Intelligent Dictation

Aryeh Stock 1, Brandon Veremis 1, Mehrvash Haghighi 1

Background: Accurate gross examination is one the most important skills of a pathologist. Clear and accurate documentation is critical for the timely and accurate diagnosis. Current gross templates provide standardization; however, they are often long, and residents have to read through the whole text to fill out the blanks. Creating and maintaining quick text dictionaries is often a labor- intensive task. We deployed PhraseExpress software as a more user-friendly alternative to current word dictation template.

Technology: PhraseExpress is an intelligent text recognition software that can create dynamic templates which alter themselves based on the details entered. Unlike other templating programs, data can be entered in a non-linear data format to complete the form. It also allows for adding basic conditional logic which improves the flexibility of data entry. It includes a cloud synchronization feature and can be shared across a network. The privacy setting can be adjusted for phrases to be private or shared with other users. This software can learn from typing behavior and includes features such as autocomplete, canned responses and macroautomation which can save hours of typing in medical terminology. It is program-agnostic and can be inserted in any word formatting software, and also can run on any operating system.

Methods: Grossing dictation templates were created in PhraseExpress for a variety of routinely accessioned specimens. Templates ranged from relatively simple to highly dynamic. Time to completion of the dynamic forms vs. regular Word templates was recorded. The number of errors in each group was assessed. A user survey was conducted to collect residents' feedback.

Results: Residents were able to complete gross dictation reports significantly faster, with fewer errors when utilizing the dynamic templates in PhraseExpress. Most residents preferred working with dynamic templates instead of navigating through lengthy documents to fill out the blanks. They also expressed more rapid acquisition of gross examination skills using the interactive format of dynamic templates.

Conclusions: Dynamic templates have the potential to speed up gross dictation while reducing errors. They also can provide guidelines to the grosser through the grossing process, alleviating the need to memorize the specific instructions for a wide variety of specimen types.

J Pathol Inform. 2019 Sep 16;10:28.

Pocket to Ceiling: Imaging Solutions to Optimize Gross Room Workflow

Hansen Lam 1, Mehrvash Haghighi 1

Background: Gross imaging of surgical specimens is paramount for the accurate gross examination and diagnosis of disease. Optimized imaging workflow can facilitate consistently high-quality gross photographs especially in high volume, metropolitan hospitals such as ours. Despite the exponential rate at which digital photography and automatic devices have developed, medical gross imaging technology lacks ergonomically well-designed hardware, intuitive software interfaces, and automation of workflow. We explore different options to address these issues and show that there are cost-effective and practical solutions.

Technology: We modeled different options using the following products: iPhone XS Max, a mobile phone, DXO One, a pocket sized camera with a 1-inch CMOS sensor that can be attached to the iPhone and iPad, Panasonic Lumix DC-ZS200, a consumer digital camera, and Arecont Vision AV20185DN, a commercial security camera. We also employed Unitwain version 2.5.4.0 (Terminal Works Ltd., Rijeka, Croatia), a software allowing automatic image import from devices to corresponding surgical cases in our pathology laboratory information system, PowerPath (version 10.2.0.91).

Methods: We evaluated three imaging systems: mobile phone with DXO One attachment, consumer digital camera, and commercial security camera. Measures used to compare each system included: installation feasibility, camera automation, workstation connection, image resolution, workflow, Unitwain compatibility, total cost, and durability.

Results: Mobile phone with DXO One attachment, despite its small size and low cost, required a simple mounting arm, granted full automation via wireless connection to mobile phone, and possessed Unitwain compatibility. However, USB connection was required after every case to import photos directly to PowerPath. The consumer digital camera allowed limited automation via USB connection to the workstation and instant image import to computer, but necessitated complex mounting systems, and carried high costs. The high-resolution security camera showed the highest level of automation in image acquisition, but required laborious installation, and carried the highest cost. Neither digital camera or security camera were Unitwain compatible.

Conclusions: Pathology gross imaging requirements are unique. Currently, there is no single option in the market that fits all our needs. Seamless workflow integration requires close collaboration of pathologists with imaging companies to customize existing devices and develop the perfect solution.

J Pathol Inform. 2019 Sep 16;10:28.

A DICOM-Based Machine Learning Workflow for Computational Pathology

Markus Daniel Herrmann 1

Background: Adoption of digital pathology is pending, and the value proposition currently hinges upon the promise that machine learning (ML) will unlock its full potential. Several ML models have shown promising results in various research settings. Now pathologists are tasked with assessing their performance in clinical practice. A major barrier for deploying and assessing models is the lack of a standard platform and workflow for clinical integration. We recently demonstrated that implementation of DICOM enables interoperability for whole slide imaging. Here, we assessed whether the DICOM standard could accommodate ML workflows.

Methods: We defined an ML model as a black-box that accepts digital images and related information as input and returns quantitative and/or qualitative predictions for those images as output. We designed and implemented a platform and workflow based on DICOM information object definitions and services to store, retrieve, and visualize model inputs and outputs. Specifically, we evaluated DICOM Structured Reporting (SR) for encoding measurements together with the relevant whole slide image regions for representation of ground truth labels and model predictions and DICOMweb RESTful endpoints and resources for model input/output.

Results: DICOM SR supports encoding of user annotations and computational measurements of whole slide images together with relevant contextual information in a structured, self- describing, and machine-readable format leveraging standard terminologies and nomenclatures that allow for unambiguous interpretation in clinical context. We developed abstractions to simplify the creation of standard SR documents in Python and JavaScript and found DICOM SR to be sufficiently flexible and extensible to support our image-related ML use cases. In addition, we built DICOMweb-based tools that enable pathologists to create annotations and visualize measurements for regions of interest in whole slide images and exchange the data in DICOM JavaScript Object Notation (JSON) format via DICOMweb RESTful services.

Conclusion: While ML models have evolved technically within the research domain, their clinical assessment represents a major operational hurdle. We present a workflow to facilitate clinical integration and proficiency testing through standard data formats, concepts, and interfaces. Our implementations are available as free and open-source software to promote data standardization in ML research and enable interoperability with clinical systems.

J Pathol Inform. 2019 Sep 16;10:28.

Query-based Digital Media Archive for Anatomic Pathology

Emilio Madrigal 1, Elliott M Moy 1, Veronica E Klepeis 1, Long P Le 1

Background: Web-based archiving and retrieval of digital media (e.g., macroscopic photography, electron microscopy [EM] and immunofluorescence [IF] images) in routine pathology practice has not achieved widespread use in part due to the technical and logical challenges of implementing capable systems. Currently, suboptimal workarounds to store and distribute such media include using shared network drives, which often limit metadata input to attributes provided in the filename. We implemented a lightweight digital media archive (DMA) to store and query data, facilitating file organization and metadata entry while reducing errors associated with manual input conforming to complex laboratory-defined naming schemes.

Methods: Within our Kubernetes cluster (container orchestration system) we built a MediaWiki stack by deploying open-source Docker images for MediaWiki (version 1.31.1), web (Apache) and database (MariaBD) servers, and PHP. The DMA leverages our institution's Lightweight Directory Access Protocol for user authentication. After installing the Cargo and PageForms extensions, templates with defined schemas were written to generate database tables on-the- fly (within the browser). Forms with various input types (e.g., dropdown, radio button, text) make calls to templates and subsequently write to the database table. For one of our use cases, a user enters an order for a new specimen by providing the accession number and medical record number with a barcode scanner. Once the specimen page is generated, the user easily stores annotated images by making drag-and-drop decisions: specimen part designator, laterality, and anatomic location. At the time of upload, unique filenames are automatically generated from the concatenation of the accession number and the digital camera assigned filename; an audit trail is saved based on the logged in user. Page cases are dynamically populated to show image files with associated metadata using customizable queries based on SQL components.

Results: The current version of the DMA is composed of 2 tables for 5 different specimen types (macroscopic photos of autopsies, neuropathology, and surgical cases, EM, and IF images) and 4 additional tables to handle unique attributes for the various media types.

Conclusions: We demonstrated the ease of setting up a DMA capable of handling various media types generalizable across the pathology laboratory with image versioning and robust audit trails. We plan to create a table for the part descriptor, clinical history, and final diagnosis from our anatomic pathology laboratory information system that will allow us to enhance the queries with more case metadata.

J Pathol Inform. 2019 Sep 16;10:28.

One Million Reasons: Creating a Repository of Over 1 Million Whole Slide Images

M C Lloyd 1, D Kellough 1, T Shanks 1, T Trefethen 1, A Parwani 2

Background: With the advent of slide scanning devices, pathologists are using whole slide images (WSI) for primary diagnosis as well as conferences, tumor boards and case sharing. Furthermore, researchers have been using whole slide images to interrogate many aspects of histology with deep learning and artificial intelligence. To facilitate meaningful artificial intelligence, large data sets are required. Our group is using process excellence to maximize the number of whole slide scans which can be acquired rapidly in the world's largest throughput pathology imaging facility.

Methods: Our group created a high-throughput scanning facility which has now scanned over 1 million whole slide images in less than 18 months. The undertaking has included high-throughput scanners from multiple vendors, full-time staff supplied by Inspirata, comprehensive planning, SOPs, project governance, workflow review, communication and stakeholder engagement activities, system integrations, file storage and IT support.

Results: We have scanned 1,011,184 slides at the time of this submission. End to end the slides would extend for nearly 50 miles and could be stacked over 4,200ft tall. They occupy about 2PB of disk storage in various levels of data temperature. We have scanned over 4 years of retrospective cancer cases while imaging prospective cases. Since our go-live date we have begun a follow-up process excellence review to enhance the following processes and increase our scanning volume: lab layout (storage and labeling procedures); slide touches (>2,600/day), slide inspection and cleaning procedures (<20 sec/slide), scan error rate reduction (<1%) and QC reviews at the scanner and for the WSI (>3%). We have run two (2) Process Excellence exercises to continue to increase our throughput by using LEAN and Six Sigma methods.

Conclusions: The promise of computational pathology is growing. It requires large data sets. By harnessing the value of a quantifiable digital pathology data modality through deep learning and similar computational advancements, it is being proven that researchers can uncover morphological signatures occurring within tumors and their surrounding microenvironment to provide better clinically actionable insights. Our high-throughput facility is the first to generate a massive pipeline of image data has not previously been achieved at this scale.

J Pathol Inform. 2019 Sep 16;10:28.

A Bunch of Barcodes for Identification of Histopathology Images

Shivam Kalra 1, Charles Choi 2, Wafik Moussa 2, Liron Pantanowitz 3, Hamid Tizhoosh 1

Background: Content-based image retrieval deals with the identification of digital images using pixel values and their features. Digital pathology can benefit from image search in large archives of whole slide images as a dynamic and smart platform to exploit the information stored in evidently diagnosed cases.

Technology: The image search is based on a combination of supervised and unsupervised algorithms. Deep and handcrafted features are employed to characterize images. The search technology is inherently “unsupervised” as it works with raw data with no specific training for the search task.

Methods: We use a cohort of different algorithms including segmentation and clustering algorithms, deep networks and distance metrics for search and retrieval.

Results: The proposed image search platform is qualitatively tested with 300 WSIs. We achieved approximately 88% accuracy in predicting the correct class for a given query image using only the first search result (the best match). We also report validation results by the pathologist that resulted in 76% accuracy for the first match of the search. Additional statistics will be provided as well.

Conclusion: The initial results of this internal validation are quite encouraging. The visual similarity of retrieved cases is for most queries striking. We continue to improve the accuracy and the speed requirements of the search platform to make it feasible and useful for diagnostic, research and educational purposes.

J Pathol Inform. 2019 Sep 16;10:28.

Characterization and Assessment of Deep Learning Systems for Histopathology Whole Slide Imaging

Weijie Chen 1, Weizhe Li 1, Brandon Gallas 1

Background: Development of deep learning (DL) algorithms for cancer diagnosis on histopathology whole slide images (WSI) is an active area of research. However, consensus is needed on approaches to characterize and assess such systems to assure quality and good science, and ultimately facilitate the deployment of such systems to the clinic.

Technology* (optional): A deep learning algorithm usually has a complex architecture involving tens of millions of parameters. Its training, validation, and testing for cancer diagnosis in pathology relies on sufficient amount of WSI data with reliable annotations of abnormalities by expert pathologists. Color image processing technologies are usually used to augment the training data. Such algorithms are expected to be robust to variations from different scanners, tissue preparation and staining protocols, and other sources.

Methods: We surveyed relevant literature regarding algorithm characteristics and assessment methods. In the meantime, we implemented a deep learning system using the Camelyon16 dataset - WSI data for breast cancer nodal metastasis detection. Based upon these, we focus our investigations on three specific topics Reproducibility: identification of critical elements in algorithm descriptions such that the implementation is reproducible (if other researchers wish to) - a basic scientific requirement; Quality and Robustness: measures taken to assure quality of data, especially the reliability of ground truth; techniques to enhance robustness to variations in image acquisition procedures; (Pre-)Clinical assessment: types of DL output, interpretability and performance metrics; characterization of false positive and false negative errors by the algorithm.

Results: Not all published work in this area provided sufficient description to make their work reproducible. We have identified critical elements that should be included in algorithm descriptions. Color augmentation and normalization techniques have been shown to be useful in improving the performance of DL algorithms, although techniques vary, and further assessment would be useful. Assessment studies are mostly pre-clinical with both slide-based metrics and location-specific metrics. Further investigations are needed on appropriate clinical use, performance metrics, and design and execution of clinical studies for such DL algorithms.

Conclusions: Consensus development on technical characterization and clinical assessment methods will be very useful in translating DL technology to clinical use for cancer diagnosis.

J Pathol Inform. 2019 Sep 16;10:28.

Effect of Feature Information-Aided Review on Pathology Trainee Performance for Ovarian Cancer Subtyping: An Observer Study

Marios A Gavrielides 1, Meghan Miller 1,2, Ian S Hagemann 2,3, Heba Abdelal 2, Zahra Alipour 2, Jie-Fu Chen 2, Behzad Salari 2, Lulu Sun 2, Huifang Zhou 2, Jeffrey Seidman 4

Background: Due to the potential for subtype-specific treatment of ovarian cancer, it is increasingly important for subtype to be diagnosed accurately and reproducibly. Pathologist performance for this task depends on the pathologist's expertise. Informatics and computer- aided diagnosis tools may be useful in narrowing this knowledge gap. This study examines the effect of providing additional information regarding histological feature presence on the performance of pathology trainees for ovarian subtyping.

Technology* (optional): Histological review was conducted on whole-slide images (WSI) using a single, calibrated monitor, and two different review modes: a) unaided, consisting of typical review of WSI images, and b) feature information-aided, or aided [Figure 1], where in addition to reviewing WSI, observers were informed about which histologic features among {high-grade nuclear atypia, abundant mitoses, intra-cytoplasmic mucin, hyalinized stroma, clear cell architectural patterns, sarcomatous components, squamous differentiation, and endometriosis}, were identified previously by an expert in gynecological pathology on the same section.

Figure 1.

Figure 1

Open in a new tab

Aided review mode

Methods: Ninety WSI from 75 ovarian cancer patients were reviewed by six 2nd and 3rd year pathology residents using unaided and informatics-aided review modes with > 3-week washout period and order re-randomization between reviews. The reference standard on ovarian subtype consisted of majority consensus from a panel of 3 experts reading on a microscope. Concordance analysis was conducted between observers and the reference standard, across the unaided and aided review modes.

Results: Aided review improved pairwise concordance with the reference standard for five of six observers, by 3.3% to 17.8% (for 2 observers, increase was statistically significant). One observer had reduced concordance by 8.9%. Difference in concordance rate between aided and unaided reviews across subtypes was [+6.9%, +2.2%, +5.6%, +8.9%, -4.4%, +23.1%] for {high grade serous, low grade serous, mucinous, clear cell, endometrioid, and carcinosarcomas} respectively.

Conclusions: Findings show the potential of information-aided review, focusing on the presence of pertinent histologic features, to improve concordance between trainees and expert pathologists for primary diagnosis tasks such as ovarian subtype classification. The observed improvement varied across trainees and histological subtypes. Future work will focus on computer-aided detection of histologic features in support of tumor classification.

J Pathol Inform. 2019 Sep 16;10:28.

Initial Clinical Validation of Clearing Histology with Multiphoton Microscopy (CHiMP) for Prostate Biopsy Diagnosis

Richard Torrez 1, Eben Olson 1, Sudhir Perincheri 2, Robert Homer 2, Michael J Levene 3, Darryl Martin 4, Preston Sprenkle 4, Peter Humphrey 2

Background: We have previously described the use of CHiMP, a direct-to-digital methodology capable of fast, easy, and cost-effective biopsy tissue processing and multi-level digital slide preparation without the need for the most technically challenging and time-consuming aspects of standard physical slide preparation. In this study we describe the first clinical validation study using this technique as applied to pathologist reads of human prostate tissue biopsies.

Methods: Single core biopsies from 20 consented individuals undergoing prostate biopsy for suspicion of prostate cancer were submitted for CHiMP imaging prior to standard histology processing. Three pathologists reviewed digital images for detailed diagnostic evaluation of all cases using web-based software, and separately reviewed physical H&E slides post CHiMP. A minimum 4-week washout period was used between review phases.

Technology: CHiMP employs a modified tissue processing protocol that includes formalin fixation with fluorescent staining during dehydration followed by optical clearing with optically matched clearing reagents for multiphoton microscopy, known as Clearing Histology with Multiphoton Microscopy (CHiMP). Most recently, we have described a specialized multiphoton microscope capable of imaging speeds comparable to whole slide imaging with high resolution and quality at depth without the need for wax embedding. Our customized web-based software for multi-level digital CHiMP images used for this evaluation is known as Stackstreamer.

Results: Specimens were received, processed, and imaged immediately post acquisition, enabling digital slide review within 3-6 hours post collection, much earlier than physical slides for the same cases were available. Concordance rates for diagnosis between digital slides and physical slides were similar to concordance rates between pathologists for physical slides alone or digital slides alone. No detrimental effects on physical slide preparation or evaluation were noted with the alternate pre-processing.

Conclusions: No significant limitations for primary diagnosis using CHiMP with prostate biopsy specimens were identified in this initial validation. There were no identifiable risks for subsequent physical slide prostate biopsy diagnosis in prostate biopsy samples previously processed by CHiMP. Next phase validation can proceed for clinical implementation of CHiMP for prostate diagnosis.

J Pathol Inform. 2019 Sep 16;10:28.

Assessment of HER2 Amplification in Invasive Breast Cancer from Chromogenic in Situ Hybridization Using Digital and Computational Pathology

Hossain Md Shakhawat 1,2, Tomoya Nakamura 1, Matthew Hanna 2, Noahiro Uraoka 2, Dara S Ross 2, Meera R Hameed 2, Masahiro Yamaguchi 1, Yukako Yagi 2

Background: HER2 gene amplification is seen in up to 20% of breast cancer and has prognostic and therapeutic indications. Fluorescent in situ hybridization (FISH) and chromogenic in situ hybridization (CISH) are the standard assays to determine the HER2 amplification status, the latter utilizing a bright-field microscope. CISH is evaluated by counting at least 20 cancer nuclei manually according to the American Society of Clinical Oncology (ASCO)/College of American Pathologists (CAP) guidelines. However, this process is time prohibitive. The purpose of this study is to develop an automated system to quantify the HER2 amplification status by CISH whole slide images (WSI), utilizing digital image analysis techniques.

Technology: A model was created to mirror the ASCO/CAP HER2 guidelines. It detected singular (non-overlapping) nuclei and identified HER2 and chromosome enumeration probe 17 (CEP17) signals per nuclei from the annotated regions. The method utilized color unmixing and machine learning techniques for nuclei detection. HER2 and CEP17 signals were detected based on RGB intensity and counted for each nucleus where CEP17≥2 and HER2>CEP17.

Methods: Patient specimens diagnosed with invasive breast carcinoma with prior immunohistochemistry (IHC) and FISH analysis were randomly selected. Subsequent manual assessment of CISH was performed. CISH whole slide images were generated at 40x (0.13 um/pixel) by the P250 3DHistech scanner. Subspecialty breast pathologists annotated regions containing invasive tumor cells. Then, the developed model quantified 20 nuclei with the highest differentiation values (HER2-CEP17). Finally, HER2 status was determined based on the HER2/CEP17 ratio as

graphic file with name JPI-10-28-g028.jpg

Additionally, another 20 nuclei were quantified if the ratio was =>1.8 and <=2.2.

Results: The proposed method was compared with manual CISH counting in terms of HER2/CEP17 ratio for 13 cases. The correlation coefficient was 0.97, which indicates the efficacy of the proposed method to quantify HER2 amplification automatically. Table 1 (next page) shows the HER2 status of 9 positive and 4 negative cases by IHC, FISH, manual CISH, and automatic CISH quantification.

Table 1.

Results for 13 cases in immunohistochemistry, flouorescent in situ hybridization, manual chromogenic in situ hybridization and proposed automatic chromogenic in situ hybridization quantification

Case IHC FISH ratio Manual CISH ratio Automatic CISH ratio
1 1-2+ 2.1 2.68 3.17
2 2+ 6.6 6.39 6.46
3 1-2+ 1 0.89 0.84
4 2-3+ 5.1 6.78 5.91
5 3+ N/A 2.82 2.41
6 N/A 2.3 2.29 2.35
7 1-2+ 2.1 1.97 1.72
8 1-2+ 1.4 1.41 1.45
9 2+ 1.4 1.22 1.79
10 2+ 2.3 1.96 1.61
11 2+ 1.1 1.11 0.63
12 3+ 2.6 2.71 2.57
13 2+ 3.1 3.13 3.62
Open in a new tab

N/A: Not available, ratio: HER2/CEP17. CISH: Chromogenic in situ hybridization, IHC: Immunohistochemistry

Conclusions: The proposed methodology has a high concordance with manual quantification. In the future, cancer regions will be detected automatically using deep learning. The final system will enable automatic cancer detection followed by the automatic quantification of HER2 amplification.

J Pathol Inform. 2019 Sep 16;10:28.

Automatic Mycobacterium Tuberculosis Detection Using Simple Image Processing with Artificial Intelligence

Hsiang-Sheng Wang 1, Wen-Yih Liang 1

Background: Tuberculosis(TB) infection is a major public health issue is Taiwan for a long period of time. During clinical practice, pathologist is one of the most important members for diagnosis. The acid-fast stain(AFS) is a special stain for identifying pathogen like Mycobacterium. However, it is a time-consuming and exhausted work for identifying small micro-organism positive in AFS even by an experienced pathologists. Due to the special feature of TB in AFS, we generate an image filter processing then combined with convolutional neural network(CNN) to detect TB automatically in virtual slides.

Methods: We first create a six-filtered method for image processing and pick up candidate by color, size, shape, color saturation, background correction and edge. The candidate image is cropped into 40 x 40 pixel small image and piping into CNN for recognition. We use tensorflow and keras as our CNN backend. The CNN structure is generated by multiple paired layers of conv2d network and maxpooling layer following 2 dense layers. The training set contains 52 positive samples where TB are labeled by experienced pathologist. Another 50 samples (All TB PCR positive but only 22 cases are TB positive by pathologist-confirmed AFS in final report) are also collected for validation.

Results: The 50 validation cases are 22 AFS positive in pathologist's final report. Our AI take total 1~27 minutes go through each case with average around 10 minutes. AI finally picks up 47 cases positive within all 50 validation cases and return the location of AFS positive area in each slide. We re-confirmed all these area and surprisingly find that only 3 of these 47 cases are false positive. The rest of 44 cases are all true positive. There is no case diagnosed TB positive missed by AI.

Conclusions: The AI system we build can work pretty well on detecting highly possible candidate within AFS. Our result also shows that AI can assist pathologist to recognized TB even in very small amount of bacilli.

J Pathol Inform. 2019 Sep 16;10:28.

Java versus Data Analysis Expressions: A Comparison of Programming Effort

Paul Christensen 1, S Wesley Long 1

Background: Analytics tools include spreadsheets with formulas, databases with aggregation functions, programming languages, and vendor software solutions. Each tool varies in its ability to support the spectrum of novice to highly-technical users in basic and complex analyses. The aim of this work was to quantitate the programming effort to perform an identical analysis in Java versus DAX (library of Excel Power Pivot functions).

Technology: Java, Microsoft Excel, Power Query, Power Pivot, DAX.

Methods: An algorithm to quantitate the impact of an HPV-reflex testing strategy from a database containing 21279 HPV tests and 59713 associated cervical biopsies was implemented in both Java and DAX. Programming time and number of lines of code were compared for both methods. Package import statements, closing curly braces and empty lines were excluded from the count.

Results: The Java solution required more programming time (2 hours versus 1 hour) and more lines of code than the DAX solution (292 versus 34 lines total, with 63 versus 16 for input/output, 181 versus 0 for the data model, and 58 versus 18 for the analysis). The DAX data model is stored in tables and requires no additional code.

Conclusions: Data analysis performed in Excel, when coupled with Power Query, Power Pivot and DAX, required less programmatic effort than the Java solution. Microsoft Office is ubiquitous in the United States workplace, and most pathologists have access to Excel. The learning curve to create Calculated Columns and Measures in DAX is lower for the average pathology resident when compared with learning how to write programming code. DAX is capable of handling complex analyses, including data with one-to-many relationships, time intelligence calculations, and analyses requiring local minimums or maximums grouped on common fields (such as a patient identifiers). Shortcomings of DAX include no built-in statistical tests or machine-learning features. Although programming languages such as Java, Python, and R are highly expressive for complicated algorithms, Microsoft Excel and its extensions may be one of the most accessible and useful data analysis tools for pathologists to learn. Training pathologists with these tools saves time in calculations and reduces manual mistakes.

J Pathol Inform. 2019 Sep 16;10:28.

From Patch-Level into Pixel-Level Annotation: Semantic Segmentation of Whole Slide Images by Histological Tissue Type

Mahdi S Hosseini 1,2, Lyndon Chan 1, Corwyn Rowsell 3,4, Konstantinos N Plataniotis 1, Savvas Damaskinos 3

Content: The gold standard for digitized histopathological diagnosis is to first identify relevant tissues in Whole Slide Images (WSI) known as regions of interest, or ROIs, and then examine these to diagnose for disease. Normally, developing an assistive computational tool to segment the WSI by tissue type would require pixel-level annotation of histological tissue type (HTT). However, pixel-level annotation is impractical considering the huge time cost and effort required by pathologists. We have sought to address this problem by training an automated segmentation tool from patch-level annotations to predict HTTs at the pixel level, assisting pathologists in identifying ROIs for diagnosis.

Technology: The Atlas of Digital Pathology (ADP) database is used to develop a weakly-supervised semantic segmentation pipeline. The ADP database is comprised of 17,688 labeled image patches extracted from 100 tissue slides (mainly H&E) across different body organs and digitized using Huron TissueScope LE1.2 WSI scanner at 40X magnification (0.25um/pixel resolution). Each patch is 1088x1088 pixels and is labeled with multi-class of Histological Tissue Types (HTT).

Methods: The semantic segmentation tool is composed of four main steps as follows. First, a patch-level HTT classifier is developed by training a Convolutional Neural Network (CNN) on the ADP database. A pixel- level HTT segmentation utility is developed next using the Gradient-weighted Class Activation Map (Grad-CAM) technique based on the class prediction. Proper adjustments are applied on the Grad-CAM features to address different characteristics of the morphological and functional HTTs. Finally, a post- processing technique is adopted to enhance the Grad-CAM segmentations to conform to the tissue contours.

Results: Figure 1 demonstrates the semantic segmentation pipeline discussed above. An image patch example of exocrine glands is fed into the pipeline to segment both morphological (predominantly loose connective, simple columnar, and leukocytes) and functional tissues (predominantly exocrine glands and transport vessels).

Figure 1.

Figure 1

Open in a new tab

Semantic segmentation pipeline based on patch-level histological tissue types classification of atlas of digital pathology database

Conclusions: The pixel-level segmentation of a WSI by histological tissue type can be achieved by training a weakly- supervised semantic segmentation algorithm on the ADP database that is labeled on the patch-level. The results suggest that the segmentation utility can be adopted by pathologists as an assistive tool to identify different ROIs on a WSI.

J Pathol Inform. 2019 Sep 16;10:28.

Digitally Tracking Manual Microscopy Slide Reading for Digital Workflow Development

J L Kohan 1, N Lobell 1, B Mathison 1, M R Couturier 1,2, O Ardon 1,2

Background: The development of new digital workflows requires an understanding of current proven laboratory workflows. Microscopy usage metrics including slide reading time, technique, and area covered by technologists are needed in order to capture the standard manual workflow to establish a comparable digital workflow. The novel development of a deep learning screening tool for a high volume, high complexity infectious disease lab test initiated a quantitative and qualitative study to capture the current manual workflow.

Methods: A video camera placed in close proximity to a clinical lab microscope filmed eight different technologists' hand movements during typical daily runs of 30 slides. The hand movement footage was analyzed using Adobe Premier for read time and number of fields of view examined on each slide. An augmented reality device capable of tracking microscope stage movements was then used as a comparison to the video recordings study. The device generated additional data including the time spent in each field of view, accurate spatial locations on the slide, and measurements of reading time spent per slide. Data was analyzed and compared to perceived microscope usage as well as video camera captured data.

Results: The mean time analyzed using the video camera recordings varied greatly between individual technologists, ranging from 26.4 - 147.5 seconds/slide. Similarly, the mean fields of view ranged from 45-196.3/slide. Additional data obtained from the augmented reality device illustrated significant variability between technologists' unique reading patterns and area covered on a slide. The results were used to determine metrics for a new machine learning tool as well as training in the lab.

Conclusions: Slide reading techniques vary greatly across different microscopy users in the infectious disease setting. These metrics are useful for machine learning based diagnostic tool development. The methods developed will be used for additional studies in different areas of the lab as digital workflows are introduced.

J Pathol Inform. 2019 Sep 16;10:28.

Improving Medical Students' Understanding of Pediatric Diseases Through Philips Pathology Tutor (formerly PathXL)

Cathy P Chen 1, Bradley Clifford 2, Matthew O'Leary 2, Douglas J Hartman 1,3, Jennifer Picarsic 1,4

Background: Online case-based modules have been increasingly integrated into medical education to optimize learning. Medical students may have access to online clinical-based modules during their clinical rotations, but they often lack information about the histopathology correlates of diseases and minimal time is devoted to pathology teaching. To address this gap, we created histopathology case-based e-modules to complement the pediatric clerkship curriculum in order to enhance medical students' understanding of pediatric diseases.

Technology: Tutor (Philips Pathology, Amsterdam, Netherlands), formerly PathXL, is an interactive web-based program licensed through the University of Pittsburgh School of Medicine Department of Pathology.

Methods: Five histopathology case-based e-modules with pre/post-tests were developed in Tutor. Each module contains a clinical vignette, digital microscopy with detailed explanation of disease process, a supplementary image, and links to additional resources. Slide annotations direct users to areas of interest. Five pre/post-test questions related to the modules were also developed in Tutor. Topics were selected based on established learning objectives for pediatric clerkships. Pre/Post-tests were administered at the beginning and end of each rotation. Test group had access to the Tutor modules, whereas control group did not. Both groups completed the pre/post-tests. Post-test was followed by a voluntary feedback survey.

Results: Twenty-two students completed the study (control group n=9, test group n=13). Test group improved their post-test scores from their pre-test scores by about one point; control group did not [Table 1]. Students responded that test questions were helpful in assessing their knowledge of the pediatric pathology (90%) and expressed relative ease of use with the Tutor program (80%).

Table 1.

Comparison of mean pretest and posttest scores

Pretest Posttest Differences Pa
Control group 3.78 3.78 0.00 1.00
Test group 3.38 4.31 0.92 0.01
Open in a new tab

aP values determined by using paired t-test

Conclusions: Medical students with access to the Tutor modules had improved post-test scores compared to those without access to the modules. Students responded favorably to the new technology. Incorporating histopathology case-based online modules into the pediatric clerkship curriculum may heighten medical students' understanding of important pediatric diseases. Our model may serve as a pilot for introduction into other medical education platforms.

J Pathol Inform. 2019 Sep 16;10:28.

Analysis of Free-Text Comments Made by Pathologists in Cancer Synoptic Reports

Veronica Klepeis 1, Emilio Madrigal 1

Background: The benefits of presenting cancer diagnoses in a synoptic report are well-recognized. However, an often- cited drawback of electronic synoptic reporting is the rigid format. A modified version of the College of American Pathologists electronic cancer checklists was implemented at our institution in which more free-text entry comment fields were included to help alleviate this issue. The goal of this study was to review all free-text entries in signed-out synoptic reports to reveal potential deficiencies and ultimately optimize synoptic templates.

Methods: All cases signed out in our laboratory information system, Sunquest CoPath Plus, between January 1, 2017 and December 31, 2018 that contained a synoptic report entered using the integrated mTuitive xPert software were exported for analysis. Each synoptic report was categorized by organ/location type. We initially focused on one of the most commonly used synoptics, namely invasive breast carcinoma (IBC). For each IBC synoptic, all free-text entries were noted and categorized by type.

Results: A total of 6501 synoptic reports were signed out over the two-year period and, of those, 23% (1524) were IBC synoptics. 68% of IBC synoptics contained one or more free-text entries. More specifically, 7% of IBC reports contained free-text entries in the header section, 55% in the invasive carcinoma section, 15% in the DCIS section, 22% in the margin sections, 13% in the lymph node section and 24% in the ancillary studies section. In the tumor section, comments on tumor size (17%), tumor grade (11%) and tumor focality (8%) were most common, followed by lymphovascular invasion (6%). Location of DCIS often required free-text comment (10%), while a lymph node comment field was used 10% of the time. In contrast, a general comment section at the end of the report was only used in 0.5% of cases.

Conclusions: Free-text entry fields in synoptic reports provide a place for pathologists to explain complexities and unique features of cancer diagnoses. Furthermore, by monitoring types of comments, synoptic templates can be optimized to include more relevant answer choices with the goal of streamlining data entry.

J Pathol Inform. 2019 Sep 16;10:28.

Canary - A Natural Language Processing Platform for Clinicians and Researchers

Alexander Turchin 1

Background: A large fraction of information in Pathology medical records is contained in narrative documents, as many nuances of tissue characterization cannot be reflected in coding terminologies. While historically narrative data were analyzed using manual chart review, recently natural language processing (NLP) technology has been increasingly used to obtain information from narrative medical documents. However, many NLP tools either require computer science training to use or are expensive. A low-cost solution accessible to users without software engineering expertise is therefore needed.

Technology: To meet these needs we developed Canary – an open source NLP platform that is aimed at clinicians, researchers and analysts without computer science expertise that is publicly available at http://canary.bwh.harvard.edu/.

Methods: Canary is a graphic user interface-based platform that allows users to create their own NLP tools using a combination of user-developed lexicons and detection criteria, ranging from simple to complex. Canary supports a number of advanced NLP functionalities, including: a) detection of concept-value pairs (e.g. tumor grade or TNM stage); b) identification of concepts distributed over multiple sentences (e.g. The specimen consists of an intact gallbladder. In the lumen there is an irregular shaped mass.); c) extraction of concept components that can be at varying distance from each other; and d) parallel processing. All Canary NLP tools are portable to any Canary installation; it has both a graphic user interface and a command-line interface for batch processing.

Results: Canary has been successfully used for development of a range of NLP tools, including diagnosis identification, medication management and analysis of imaging reports. Sensitivity and positive predictive value of NLP tools developed using Canary platform has ranged between 80-95% and it has reached speeds of text processing at 1 MB / minute / CPU core. A number of NLP tools developed for the Canary platform are publicly available in the Canary NLP Tool Library (http://canary.bwh.harvard.edu/library/).

Conclusions: Canary is a versatile NLP platform that can be effectively used by clinicians and researchers to extract information from electronic medical records. It is particularly promising for applications in specialties, like Pathology, where a large fraction of data is contained in narrative reports.

J Pathol Inform. 2019 Sep 16;10:28.

Cognitive Informatics

Edward Klatt 1

Background: Cognitive informatics combines cognitive, behavioral, and information sciences to inform design of health information technology through analysis of human information processing and collaborative requirements of work being done by end users. People are the ultimate users of biomedical information.

Methods: A literature review regarding human cognition informs refinements for providing information to the healthcare team. Informaticians can recognize limitations of human cognition and draw upon cognitive science to inform the design and evaluation of technical solutions for information management and interface with the healthcare team.

Results: The human brain has a finite capacity for processing new information because neuronal synapse formation is a rate-limited process, called long-term potentiation. Short-term working memory may be limited to no more than 4 separate informational items processed simultaneously. Cognitive load is reduced by breaking down complex tasks into a series of simplified tasks. Long- term memory supplies immediate access to multiple informational items simultaneously. Visual long-term memories can be extensive and detailed, so use of imaging is effective for providing information. Attention span may be task dependent and highly variable among persons. Effective attention span for learning may not exceed 20 minutes. Attention requires control over distracting information. Noise is most distracting when it more closely resembles recognizable human speech. Multitasking is multisequencing, and adding tasks leads to performing each in shorter sequences, because tasks compete for working memory, reducing the effectiveness of working memory applied to each task. Workplace stress with emotional arousal adversely impacts memory storage and function when too high.

Conclusions: Solutions applying cognitive informatics include methods to address cognitive load. Medicine is data-rich, but the data are complex. Visual analytics is a means to provide advanced interactive visual interfaces to aid reasoning over, and interpretation of, complex data to avoid information overload. Deep learning describes machine learning algorithms capable of combining many raw inputs into layers of intermediate features. Deep learning has the potential to address problems of cognitive load and constraints of human information processing. Application of deep learning to imaging informatics includes repurposing features extracted from natural images for training, then applying the model to real-world problems.

J Pathol Inform. 2019 Sep 16;10:28.

From Value Stream Mapping to Value-Based Health Care

Mehrvash Haghighi 1

Background: Health care organizations, due to the transition from fee-for-service to value-based systems, started to embrace lean methodology as a business strategy to achieve better operational efficiency and cost reduction. Educating resident competence in understanding, developing, and implementing successful QI projects is essential for their future success as a pathologist in the value-based era.

Methods: We adopted the Juran principle as the foundation of our QI curriculum with a strong emphasis on value stream mapping. Residents were divided into six groups with relatively balanced gender composition and level of training. The program contained four lectures (proof of need, project identification, diagnostic journey, and remedial journey) with associated activities/games, a case study session, and performance of a mentored QI project. In the case study session, a real business manufacturing case was presented to the residents, and they were asked to perform workflow analysis, draw value stream mapping, propose a solution, and calculate the improvement in performance indexes.

Results: The average attendance rate was 90%. The junior residents in the order of first- and second- year residents showed the highest level of engagement in weekly assignments. There were no significant differences in performance based on gender/age. The results indicated that residents had increased their knowledge in QI methodologies. Four out of six teams (66%) delivered all phases of the project performance (quality planning, control and improvement) with a perfect score. One of the completed projects addressed an issue related to the autopsy department at the organizational level which resulted in 50% improvement in performance in the early stages of implementation. For this specific project, a combination of different strategies including repair, renovation, and reinvention were deployed. In other three projects, quality improvement was achieved through advancement of technologies. One of the team submitted the project end-result as an innovation proposal for grant approval.

Conclusion: The objectives of this program included educating residents to identify a problem/defect, illustrating the issue using value stream mapping, selecting the best solution using different methods and techniques, defining a project to solve the problem. We consider the successful implementation of the QI projects to be the evidence of an effective learning mechanism for advancing resident education in quality improvement and project management.

J Pathol Inform. 2019 Sep 16;10:28.

Converting the Narrative to Analytics: Unlock the Value of Your Data

Aryeh Stock 1, Brandon Veremis 1, Mehrvash Haghighi 1

Background: Currently, surgical pathology reports are stored as a free-text variable character field. This means that a huge amount of time is spent on hunting and gathering rather than understanding and interpreting data and creating valuable prognostic metrics. The developed program can navigate the narrative diagnostic text and automatically parse it to the part type and corresponding diagnosis in an Excel (Microsoft) spreadsheet. This program opens the door to readily available quantifiable data on a wide range of clinical outcomes. Examples include the ability to track changes in the progress of disease over time, the ability to correlate various surgical techniques with an adverse event timeline (e.g., complications, severity of complication). Mount Sinai has vast amounts of clinical data that are being underutilized, simply because it is saved in a format that is not “analysis-friendly.”

Technology: Scripts were written in Google Docs and Python (3.27.2) with NumPy (1.15.4) and pandas (0.23.4) libraries.

Methods: Utilizing rule-based natural language processing and syntax recognition, scripts in Google Docs and Python were written to parse narrative surgical pathology reports and generate database-ready spreadsheets.

Results: This script can rapidly identify, sort and assign numerical values to the findings of a narrative-based pathology report. A proof-of-concept script was written that can navigate colon biopsy reports and generates a heat map of the severity of inflammation throughout the colon.

Conclusions: Short-term applications allow for the integration of heat maps on colonoscopy reports to show how the severity of IBD changes throughout a patient's colon and over time. Broader applications include the potential to extract quantifiable data from a variety of pathology reports for use in research and value- based care. This application is best-suited to divisions that use pre-formatted codes to write their reports. Future directions include incorporating statistical natural language processing to extract data from reports written with less consistent use of language.

J Pathol Inform. 2019 Sep 16;10:28.

Informatics Education in Pathology Residency Programs: A Successful Team-Based Approach

Yonah Ziemba 1, Tarush Kothari 1, Vanesa Bijol 1, Kalpana Reddy 1, Michael Esposito 1

Background: Training gaps in pathology residency curriculum have recently become well-recognized, and surveys show that newly-employed pathologists identify informatics as an area where training is insufficient for successful practice. This is partly because informatics is not amenable to the common training model in which residents rotate on clinical services to assist attendings in their daily work. Herein, we describe a successful team-based learning (TBL) approach implemented at our residency program that does not require dedicated faculty. This has been implemented for two years with remarkable success and is accessible to any program.

Methods: TBL sessions that correspond to topics in the Pathology Informatics Essential for Residents (PIER) Resource Toolkit were attended monthly by 16 residents. Preparation for the session consists of PIER's recommended resources and practice exercises. The discussion focuses on each resident's implementation of the practice exercises, or on scenarios that mirrored the PIER learning objectives. Faculty merely function as moderators, and the content is driven by the residents.

Results: To quantify effectiveness, we reviewed studies that assess TBL models similar to ours. Studies measured one of three outcomes: engagement, satisfaction, and examination scores. Kelly et al reported that students were significantly more engaged during TBL than during lecture, as measured by the STROBE tool. Balwan et al found that TBL was associated with increased satisfaction among faculty and residents when implemented at an Internal Medicine Residency Program. Koles et al reported that TBL was associated with higher scores on medical school examinations. In addition, authors note that advantages to TBLs compared to lecture include development of creativity, critical thinking and interpersonal skills.

Conclusions: Active learning in TBL is more effective than passive learning in lecture, and TBLs in pathology informatics are accessible and easy to implement. The resources provided by PIER are flexible and can be adapted to a team-based approach for a large group, or a project-based approach for a single resident on an Informatics rotation. Strengths of PIER include that it is free to all pathology residency programs, it can be adapted to fit a variety of curriculum structures and can be implemented without the benefit of dedicated informatics faculty.

Articles from Journal of Pathology Informatics are provided here courtesy of Elsevier

RESOURCES