Feasibility and preliminary accuracy of high-resolution imaging of the liver and pancreas using FNA compatible microendoscopy

Abstract

Background

Endoscopic ultrasound-guided fine needle aspiration (EUS-FNA) is one of the few techniques which can obtain cells and tissue from the liver and pancreas. However, the technique remains vulnerable to poor specimen quality and sampling error.

Objectives

To evaluate the ability of a high-resolution microendoscope (HRME) to visualize the cellular and architectural features of normal and malignant liver and pancreatic tissue ex vivo. To assess the ability of endosonographers to identify normal and neoplastic tissue using HRME images. To demonstrate preliminary technical feasibility of in vivo HRME imaging via EUS-fine needle puncture (FNP).

Design

Ex vivo, pilot feasibility study in human tissue; in vivo swine model.

Setting

Two academic medical centers

Patients and Interventions

Co-registered HRME images and biopsies were obtained from surgically resected hepatic and pancreatic tissues from a total of 44 patients. Images were divided into training (12 images) and test sets (80 images) containing a range of normal and pathologic conditions for each organ. After viewing the training sets, nine endosonographers attempted to distinguish malignant tissue from normal or benign lesions in the test sets, each of which contained 40 unique images with individual diagnoses from pathology.

Main Outcome Measurements

Image acquisition feasibility, ex vivo and in vivo. Ability of endosonographers to recognize features of normal/benign or malignant tissue from the liver and pancreas.

Results

Overall, the nine endosonographers achieved median accuracy figures of 85% in the liver and 90% in the pancreas. The endosonographers with prior experience in reading HRME images achieved accuracy rates between 90% and 95%. Technical feasibility of HRME imaging through a 19-gauge EUS-FNP needle was demonstrated in an in vivo swine model.

Limitations

Ex vivo study

Conclusions

High-resolution microendoscopy allows real-time imaging of cellular-level morphology and tissue architecture in the liver and pancreas. The techniques appears to have a short learning curve, after which endosonographers achieved high accuracy rates in distinguishing malignant tissue from normal and benign pathology in both organs. Translating this imaging platform to the in vivo setting appears technically feasible.

Introduction

Endoscopic ultrasound (EUS) is commonly used to visualize intra-abdominal organs including the liver and pancreas, and in conjunction with fine needle aspiration (EUS-FNA) to sample tissue from these organs. In early studies, EUS-FNA was found to be equal or superior to other imaging modalities including CT and MRI for assessing tumor size and lymph node involvement¹. However, EUS-FNA has been shown to have a failure rate of up to 10% due to poor specimen quality, sampling error, and its inability to distinguish confounders such as acute and chronic pancreatitis^2-5.

Background

Several real-time adjunctive techniques are being developed to improve the diagnostic accuracy of EUS-FNA, including contrast enhancement, elastography, and confocal laser endomicroscopy^6-11. However, due to factors including cost, complexity, and their associated learning curves, these methods have not yet gained widespread acceptance in clinical practice. We have recently developed a high-resolution microendoscope (HRME) which is a fiber-optic probe-based system which provides images of cellular morphology and tissue architecture in situ and in real-time (fig. 1a,b)^12,13. This prototype device costs less than $2,500 in components, with a probe that can be passed though a 22-gauge needle and be sterilized and reused (fig. 1c). Previous ex vivo and in vivo studies in the esophagus and colon demonstrated the feasibility of obtaining high-resolution images of the epithelial mucosa using the HRME, aimed at providing targeted guidance for biopsy collection^14,15. A recent study demonstrated the technical feasibility of performing confocal endomicroscopy in the pancreas in vivo via a 19-gauge FNA needle, with image quality described as “good to very good” in 10 out of 18 cases¹⁶. HRME imaging does not provide the optical sectioning ability of confocal endomicroscopy, but has nevertheless demonstrated very good image quality in other gastrointestinal organs in vivo, and is significantly less expensive than confocal platforms¹⁵. The aims of the study reported here were (1) to identify the characteristic features of normal tissue, benign lesions, and neoplasms in the liver and pancreas which are apparent under HRME imaging, and (2) to estimate the accuracy and learning curve for endosonographers to identify malignant neoplasms in these organs using images obtained with the HRME.

Figure 1.

The high-resolution microendoscope (HRME). (a) Schematic diagram. (b) Photograph of the HRME unit, measuring 10” × 8” × 2.5”. (c) Photograph of fiber optic probe with 0.45 mm diameter passed through a EUS-FNA needle (Cook ECHO-1-22).

Methods

Under an IRB approved protocol, we obtained written informed consent from 32 patients undergoing liver resection and 12 undergoing pancreatic resection. The patient population comprised 25 women and 19 men, with a mean age of 60 years. Technical details of the HRME have been described in detail elsewhere¹³. Briefly, the device essentially operates as a fluorescence microscope, coupled to a fiber-optic imaging probe (fig. 1a,b). Following application of a fluorescent contrast agent, the probe is placed in contact with the tissue, with an image of the tissue site transmitted back through the fiber-optic probe and captured by a CCD camera. In common with previous studies by our group^14,15 and others in the GI literature^9,17 we used 0.01% w/v proflavine solution as the fluorescent contrast agent to label cell nuclei. The HRME system used here to image human tissues has 4.4 μm spatial resolution, a 0.72 mm diameter field of view, and displays images in real-time at 12 frames per second. Use of a probe with 0.33 mm field of view allows passage through a 22-gauge EUS-FNA needle (Cook ECHO-1-22) (fig. 1c).

Immediately following resection, each fresh tissue specimen was evaluated by a gastrointestinal pathologist who identified up to four normal and four neoplastic sites by gross examination. Proflavine was topically applied to the tissue surface and HRME images were immediately acquired by direct placement of the probe on the tissue surface at each designated site. A single application of contrast agent was typically sufficient for several minutes of imaging; no rinsing of excess dye was required. Sample sections from each imaged site were then placed in formalin and submitted for histopathologic processing. Each specimen was given a histologic diagnosis by an expert pathologist using standard criteria, blinded to the corresponding HRME image.

To assess the ability of endosonographers to identify malignant neoplasms on HRME, individual images were assigned to separate “training” or “test” data sets for both the liver and the pancreas. The training set for the liver contained seven unique HRME images from a range of biopsy-proven normal, benign, and malignant tissue sites. The training set for the pancreas contained five HRME images. All endosonographers viewed the same training and test images in the same order. Each test set comprised 40 unique images which were not included in the corresponding training set. Examples of hepatic pathology included lesions such as cholangiocarcinoma, angiomyolipoma, and metastatic colon cancer, with pancreatic pathology including such conditions as serous microcystic cystadenoma and ductal adenocarcinoma. Nine endosonographers from multiple regional academic institutions participated in this study, each having performed more than 1000 EUS cases. The endosonographers were blinded to all patient history for the purposes of the study and were not involved in the generation of the training and test image sets. After viewing the training sets and being informed of the characteristic image features of each tissue type, participants were asked to evaluate the test sets for the both organs and state whether each image represented normal, benign, or malignant tissue.

To evaluate the technical feasibility of performing in vivo HRME imaging in conjunction with EUS-FNP, experiments were performed in a swine model. All experiments were approved by the Institutional Animal Care and Use Committee (IACUC) at The University of Texas MD Anderson Cancer Center. Following administration of anesthesia, a linear echoendoscope (Olympus GFUC30P) was introduced into the stomach. The stomach wall was punctured with a 19-gauge EUS-FNA needle (Cook, ECHO-1-19) and advanced into the pancreatic parenchyma under EUS guidance. Following removal of the needle stylus, 5 ml of proflavine was injected through the needle. The HRME imaging fiber was then passed through the needle into the pancreas for imaging.

Results

HRME Image Features

Figure 2 shows HRME images and corresponding histopathology from the liver. On HRME imaging of normal hepatic tissue (fig. 2a), individual nuclei are bright, evenly spaced, regularly shaped, and round to oval. In the corresponding H&E stained section (fig. 2d), normal hepatocytes have small, regularly spaced and centrally located round nuclei. In the HRME image of a benign angiomyolipoma (fig. 2b), there are large dark spaces representing lipid vacuoles outlined by haphazardly arranged cells with small, well defined nuclei. The corresponding H&E section (fig. 2e) shows a mix of adipocytes with large vacuoles, among chronic inflammatory cells and well differentiated smooth muscle cells with discrete nuclei. Tissue at the site of a metastatic colon adenocarcinoma shows loss of normal hepatic architecture in both the HRME image (fig. 2c) and the H&E section (fig. 2f) with poorly formed, irregular glandular structures leading to apparent nuclear crowding and clumping.

Figure 2.

Representative HRME images (a,b,c) of normal, benign, and malignant lesions of the liver with corresponding photographs of H&E histopathologic sections (d,e,f). (a,d) Normal liver. Note the regularly shaped, widely and evenly spaced nuclei. (b,e) Angiomylipoma, benign lesion. Note the adipocytes with clear vacuoles. (c,f) Metastatic colon adenocarcinoma, malignant lesion. Note the loss of normal architecture with irregular glandular structures and nuclear clumping. Scale bars represent 100 μm.

In the normal pancreas, nuclei appear as clustered small bright dots in the HRME image (fig. 3a), while the H&E section reveals small nuclei grouped in well formed, regularly spaced round acinar structures (fig. 3d). In benign microcystic adenoma, large cystic spaces of variable shapes and sizes can be identified in both the HRME image (fig. 3b) and H&E section (fig. 3e). Real-time HRME imaging at this site is shown in accompanying Video Clip 1. In ductal adenocarcinoma, there is loss of normal architecture with irregular clumps of streaking nuclei in the HRME image (fig. 3c) and, correspondingly, small irregular glands infiltrating amidst a desmoplastic stroma in the H&E section (fig. 3f). No normal acini are seen.

Figure 3.

Representative HRME images (a,b,c) of normal, benign, and malignant lesions of the pancreas with corresponding H&E histology (d,e,f). (a,d) Normal pancreas. Note the clustering of nuclei into acinar structures. (b,e) Microcystic adenoma, benign lesion. Note the cystic spaces of varying shapes and sizes. (c,f) Ductal adenocarcinoma, malignant lesion. Note the loss of normal architecture and infiltrating poorly formed glands amidst desmoplastic stroma. Scale bars represent 100 μm.

The primary focus of this study was to establish the characteristic features of normal, benign, and malignant tissue in the liver and pancreas on HRME imaging, by using an ex vivo study with well correlated histopathology sections. To make an initial assessment of the technical feasibility of translating this imaging method to the in vivo setting, we performed a small series of experiments in an in vivo swine model, allowing us to assess factors including contrast agent delivery, intraparenchymal image quality, effect of blood in the field, and subject motion artifact. Figure 4 presents a representative HRME image from the swine pancreas, acquired in vivo using EUS-FNP. As with the normal ex vivo human pancreas, nuclei appear as discrete bright dots, regularly spaced throughout the HRME field-of-view. We found that delivery of the contrast agent was straightforward, and manipulation of the HRME probe was essentially comparable to working with the EUS-FNA device alone.

Figure 4.

Representative HRME imaging of the pancreatic parenchyma in an in vivo swine model. Images are acquired with the fiber-optic probe advanced within the lumen of a 19-gauge EUS-FNA needle. Nuclei appear as small, discrete dots within the field of view. Scale bar represents 100 μm.

HRME image feature recognition and learning curve

Two of the nine endosonographers who completed the test sets had previous experience in viewing and interpreting HRME images (> 50 cases each), while seven had no prior experience in using HRME or interpreting HRME images. For the group as a whole, the median accuracy for identifying malignant versus normal or benign tissue in the liver was 85%, with 81% sensitivity and 88% specificity. The median positive predictive value (PPV) and negative predictive value (NPV) for detection of malignant lesions of the liver were 81% and 87% respectively. In the pancreas, the median accuracy of the group for identification of malignant lesions was 90%, with sensitivity and specificity 85% and 90%, respectively. The PPV and NPV for identifying pancreatic malignancies were 90% and 86%. These data are presented in fig. 5a and for each individual endosonographer in Table 1.

Figure 5.

(a) Performance of all nine endosonographers in reading HRME image test sets. Bars represent median values. Error bars indicate the interquartile range. (b) Performance of endosonographers with no prior experience in HRME imaging (“HRME novices”, n = 7) compared to endosonographers with prior experience (> 50 cases) in evaluating HRME images (“HRME experienced”, n = 2). Bars represent median values. Error bars indicate the interquartile range.

Table 1.

Performance of endosonographers in recognizing normal / benign tissue from malignant in test sets of 40 HRME images for each organ. There were 24 images with pathology considered malignant for the liver test set, and 20 in the pancreas test set. HRME “experienced” endosonographers had reviewed images from over 50 HRME cases prior to this study. HRME “novices” had no prior experience in HRME imaging prior to this study. All values are percentages, rounded to the nearest whole number.

	HRME experienced		HRME novices							Median
	1	2	1	2	3	4	5	6	7	Exp.	Nov.	All
Liver:
Accuracy	95	85	90	63	75	83	70	85	85	90	83	85
Sensitivity	100	81	81	63	38	81	50	63	88	91	63	81
Specificity	92	88	96	63	100	83	83	100	83	90	83	88
PPV	89	81	93	53	100	77	67	100	78	85	78	82
NPV	100	88	89	71	71	87	71	80	91	94	80	87
Pancreas:
Accuracy	93	98	65	55	90	65	68	90	93	95	68	90
Sensitivity	95	95	65	45	85	45	75	90	95	95	75	85
Specificity	90	100	65	65	95	85	60	90	90	95	85	90
PPV	91	100	65	56	94	75	65	90	91	95	75	90
NPV	95	95	65	54	86	61	71	90	95	95	71	86

The two endosonographers with prior experience in HRME were better at classifying images than those without, achieving accuracy figures of 90% in the liver and 95% for the pancreas, compared to median values of 83% and 68% respectively for the “novice” group. The “HRME experienced” endosonographers achieved sensitivity and specificity figures of 91% and 90% respectively for identifying malignant sites in the liver, compared to 63% and 83% respectively for the “HRME novice” group (fig. 5b). In the pancreas, the “experienced” group achieved sensitivity and specificity of 95% and 95% respectively for detecting malignant lesions, compared to 75% and 85% for the “novice” group. When “normal” and “benign” categories were grouped for analysis, Cohen's kappa inter-rater reliability for the two “HRME experienced” endosonographers was 0.80 for both the liver and the pancreas. Fleiss’ kappa for the group of seven “HRME novice” endosonographers was 0.39 for the liver and 0.27 for the pancreas. When all three categories (normal, benign, malignant) remained separate, Cohen's kappa for the experienced group was 0.70 for the liver and 0.84 for the pancreas. Fleiss’ kappa for the novice group was 0.44 for the liver and 0.22 for the pancreas.

Discussion

Several endoscopic techniques for evaluating tissue structure and cell-level morphology have been developed recently, the majority focusing on applications in the esophagus and colon^9,18-20. A feasibility study demonstrating confocal endomicroscopy of the pancreas was recently reported, highlighting the potential for endomicroscopy to serve as an adjunct to EUS-FNA¹⁶. While histopathologic evaluation will remain the gold standard for diagnosis, the advantage of high-resolution endomicroscopy is that real biopsies can be performed in a more targeted manner, potentially increasing diagnostic accuracy. The ability to obtain real-time microscopic information during EUS-FNA would also be particularly helpful where on-site cytopathologic evaluation is not available. Despite these potential benefits, two key factors may influence the degree to which these technologies are adopted into clinical practice. The first concerns the accuracy of the technique and the associated learning curve for the anticipated users. A second factor which impacts the diffusion of any new technology is instrument cost. We began to address these questions here by developing a low-cost instrument with a re-usable fiber-optic probe ($2,500 processor, $500 probe), and measuring the ability of endosonographers to differentiate neoplastic lesions in both the liver and pancreas.

Our group of 9 endosonographers identified malignant lesions of the liver and pancreas in 40-image test sets with a median accuracies of 85% and 90% respectively, despite the fact that 7 participants had no prior experience with microendoscopic imaging. The 2 experienced endosonographers with over 50 cases each involving HRME imaging achieved an average accuracy of 90% and 95% for the liver and pancreas, respectively, suggesting that interpretation of HRME images has a relatively short learning curve. A small number of studies have attempted to evaluate the learning curve for endomicroscopy in other organs. Buchner et al. found that users of probe-based confocal endomicroscopy were able to interpret images for identification of colorectal polyps with an accuracy of 93% after reading at least 35 cases²¹. However, it is difficult to compare post-training accuracy figures for different organ sites, and also for systems using different contrast agents which require recognition of different classification features. Nevertheless, it appears that confocal endomicroscopy and HRME may have similar learning curves, on the order of a few 10's of cases. In addition to this purely qualitative interpretation of images, there is clearly scope for quantitative image analysis to measure morphological characteristics such as total number of nuclei per field, average nuclear size, and nuclear–to-cytoplasmic ratio. Values from the most diagnostically relevant parameters can be used to create an automated algorithm to provide an objective evaluation of tissue in real-time during the imaging process.

The primary limitation of this study was the use of ex vivo surgical specimens. Our feasibility experiments in an in vivo swine model permitted assessment of in vivo factors including motion artifacts and sampling error, and the impact of bleeding on image acquisition and interpretation. The experience gained with this model suggests that in vivo deployment in humans is technically feasible. The current study design enabled us to use co-registered images and corresponding tissue sections to provide an accurate pathologic diagnosis as the gold-standard, which was not the case in the in vivo study reported by Konda et al. ¹⁶. We characterized image features in normal, benign, and malignant tissues of both the liver and pancreas, and assembled nine endosonographers with varying levels of experience to establish both the accuracy and estimated learning curve for HRME. Recognition of benign conditions including angiolipoma and serous microcystic adenoma could be more difficult in practice, due to the fact that characteristic voids in the nuclear staining pattern on HRME could also arise from other disorders, both benign and malignant, such as angiosarcoma, peliosis hepatitis and NASH. An additional limitation was related to the availability of multiple tissue specimens with a wider range of pathology, such as intraparenchymal sites, which would have enabled us to train and test the endosonographers more comprehensively. A smaller fiber-probe with correspondingly reduced field-of-view will need to be used for passage through a 22-gauge EUS-FNA needle; however, the resolution (and therefore image quality) is not affected by the fiber diameter and is expected to be maintained at the level shown here.

In order to translate this technology into the patient setting several issues will need to be addressed. Delivery of the fluorescent contrast agent to the target organ prior to imaging could be performed by first delivering the dye through the EUS needle itself, followed by insertion of the HRME fiber. This was the approach we adopted for the in vivo swine imaging. Alternatively, Tanbakuchi et al. demonstrated a clinical confocal system with a contrast agent delivery channel mounted directly alongside the imaging fiber bundle²²; such an approach could be adapted for HRME imaging within an EUS needle. These approaches could be applied to topical agents, however, as reported by other groups in the confocal endomicroscopy literature^16,23, intravital imaging dyes such as fluoroscein or indocyanine green can be delivered by an intravenous route. The fluorescent dye used here (proflavine) has been used by our group and others for several clinical endomicroscopy studies^9,15,17, however it does not currently have full FDA approval as an imaging agent. Our experience with proflavine, used under Investigational New Drug (IND) status in the esophagus and colon since 2009 has not resulted in any adverse events and we are continuing to monitor subjects enrolled in these studies.

A second intriguing question for future clinical use is how to best integrate image information from high-resolution microendoscopy with that from the EUS imaging modality. It seems likely that the endosonographer would wish to integrate macroscopic information from EUS with the microscopic scale information from HRME, and future prospective study of EUS/FNA cases will be designed to take this into account.

In conclusion, we demonstrated that a novel, low-cost endomicrocopy device could be used to acquire high-quality, high-resolution images of cellular morphology and architecture in surgical specimens from the liver and pancreas. In vivo translation appears feasible based upon pilot animal data. With experience in HRME image interpretation, users achieved diagnostic accuracy rates of up to 95%. This innovative technique can potentially improve further upon EUS-FNA, increasing its diagnostic accuracy with minimal added cost and risk.