Enhancing the Value of Histopathological Assessment of Allograft Biopsy Monitoring

Abstract

Introduction:

Traditional histopathological allograft biopsy evaluation provides, within hours, diagnoses, prognostic information, and mechanistic insights into disease processes. However, proponents of an array of alternative monitoring platforms, broadly classified as “invasive” or “noninvasive” depending on whether allograft tissue is needed, question the value proposition of tissue histopathology. The authors explore the pros and cons of current analytical methods relative to the value of traditional and next generation histopathological evaluation of tissue biopsies.

Objectives:

Illustrate the continuing value of traditional histopathological tissue assessment and “next generation pathology (NGP)”, broadly defined as staining/labeling techniques coupled with digital imaging and automated image analysis.

Summary:

Noninvasive imaging and fluid (blood, urine) analyses promote low-risk, global organ assessment, and “molecular” data output, respectively; invasive alternatives promote objective, “mechanistic” insights by creating gene lists with variably increased/decreased expression compared to steady-state/baseline. Proponents of alternative approaches contrast their preferred methods with traditional histopathology that: 1) fail to cite the main value of traditional and NGP - retention of spatial and inferred temporal context available for innumerable objective analyses; and 2) belie an unfamiliarity with the impact of advances in imaging and software-guided analytics on emerging histopathology practices. Illustrative NGP examples demonstrate the value of multi-dimensional data that preserves tissue-based spatial and temporal contexts.

Conclusion:

We outline a path forward for clinical NGP implementation where “software-assisted sign-out” will enable pathologists to conduct objective analyses that can be incorporated into their final reports and improve patient care.

INTRODUCTION

Strengths of Traditional Histopathology and Modern Alternatives

Traditional histopathology strengths and value derive primarily from its visual nature and retention of spatial and inferred temporal context at a submicron resolution level. Up to 80% of environmental stimuli are processed via visual pathways¹ and ~30%–50% of the cerebral cortex is devoted to processing visual signals.² When pathologists view organs at submicron resolution levels, hardwired visual dominance pathways immediately integrate inputs into legacy networks of pathophysiological insights. Pathologists can rapidly, reliably, and accurately: 1) identify rare (eg, isolated intimal arteritis, viral inclusions) and/or focal (eg, scars, lymphoid nodules, etc) events; 2) appreciate social relationships among cells (eg, lymphocytic tubulitis, capillaritis); 3) “age” tissue alterations into acute (hours to days), subacute (days to weeks), and chronic (months to years); and 4) recognize and rank order multiple insults.

Rapid conceptualization by humans/pathologists, comes at the expense of more precise quantitative abilities³ leading to important criticisms of histopathology: conventional tissue staining combined with human-only visual analysis is semiquantitative, thus, somewhat subjective, non-mechanistic and poorly reproducible.^4–8 Humans are intrinsically built for estimation and not precise counting.³ Other shortcomings of traditional histopathology may include: sampling errors, biopsies are invasive, risky, and an unpleasant patient experience; and granular morphological or contextual information might not be required in some circumstances.

Considerations regarding the value of liver, kidney, or heart allograft biopsy acquisition for histopathological examination are evolving. The English Oxford Dictionary defines “value” as “the regard that something is held to deserve; the importance, worth, or usefulness of something.”⁹ Under noncomplex clinical circumstances (eg, slowly evolving presumed single insult), the value of traditional histopathological biopsy analysis is questioned allowing newer noninvasive imaging tools and high throughput tissue, blood, and bodily fluid -omics platforms combined with software-assisted analytics to enter the diagnostic tool kit.

Advocates of newer noninvasive, imaging (reviewed in references 10-14) and fluid (eg, urine or blood^6,15–19) analytical techniques promote safety and global organ assessment, mitigating biopsy shortcomings of riskiness, sampling errors, and undesirability of use for serial measurements.^{5,8,10,13,20,21} However, noninvasive radiographic approaches: 1) produce low spatial resolution output; 2) often rely on indirect measurements (eg, liver stiffness as a surrogate of fibrosis) interfered with by non-fibrosis-related processes (eg, intraorgan congestion, inflammation, and patient obesity); and/or 3) introduce operator-dependent measurement variations (reviewed in references 10-14). Blood and fluid analyses provide no spatial context regarding intra-allograft events and only indirectly monitor allografts.

Advocates of high-throughput alternative tissue methodologies (eg, -omics platforms) promote value over traditional histopathology with respect to objectivity, data output, and “mechanistic” insights. Again, valuable spatial and inferred temporal contexts are destroyed while risks, patient avoidances, and sampling issues are the same as traditional biopsy analyses. Moreover, these methods use traditional histopathology for validation and often yield a narrow range of simplified single-insult diagnoses,^{4,7,17,18,22–25} that may overlap (eg, BK nephropathy and TCMR).^7,17,23–25

Complex data outputs are visually displayed for human consumption (eg, scatter plots, heatmaps, volcano plots) or reported as “gene lists” without rank ordering insults when more than one insult is present. Gene lists reported can vary by organ, circumstance, or lab.¹⁷ Sample preparation and storage, interlaboratory assay reproducibility, and interpretive expertise distribution needs standardization, resulting in calls for centralization.^7,26 Large transplant centers will evaluate the cost-benefit ratio of waiting several days (mRNA expression arrays) vs. the relatively short (hours) turnaround time of biopsy evaluation and the ability to confidently act on results for patient management.

Nevertheless, “gene lists” derived from -omics platforms represent an advancement and provide important details that can be layered on legacy histopathology mechanistic insights.^27–41 For example: 1) transcripts during TCMR show increased leukocyte populations (eg, T, B, and NK cells, macrophages) and monocyte-macrophage-derived chemokines^17,23,33,42; 2) endothelial cell mRNA profiles are perturbed in AMR^{25,28,31,42–44}; and 3) chronic inflammation, mast cell, and immunoglobulin transcripts increase over time in kidney allografts.^7,45 How effectively non-histopathology-based platforms will adapt to clinical care complexities remains forthcoming. Will “grind and bind” -omics platforms be able to: 1) rank multiple insults, 2) recognize focal defects and rare events, 3) adapt to variations in tissue age and sex, 4) time post-transplant, 5) inform drug regimens, 6) provide location data regarding donor MHC antigen expression and 7) provide spatial interactions among various cell types?

Interstitial inflammation scoring in kidney allografts encompasses peculiarities not encountered with liver allografts. If renal interstitial inflammation is located in areas of scarring or occupies <10% of non-scarred renal cortex, a Banff “i” (inflammation) score of “i0” (no inflammation) is registered.⁴⁶ Any liver inflammation is registered as abnormal. “Non-scored” (i0) kidney lymphocytic infiltrates are undoubtedly contributing to the mRNA profile, which questions the value of correlations between kidney histopathology and mRNA expression profiles.

Roles of Traditional Histopathology

Traditional histopathological liver biopsy evaluation (within hours) provides^47,48: 1) diagnosis(es); 2) pathophysiological processes involved in tissue injury (eg, TCMR, AMR, vascular issues); 3) a rank ordering of multiple simultaneous insults; 4) structural/architectural integrity assessment, including potentially irreversible damage; 5) guidance for therapeutic intervention; and 6) potentially excess tissue for translational research purposes. These traditional roles are based on hundreds of thousands (if not millions) of assays, validated over decades, and are widely available because of distributed expertise, and used as endpoints in clinical trials,^49–53 and included in USFDA guidelines.⁵⁴

Ranking multiple insults and recognizing mechanisms of tissue injury are increasingly important as noninvasive radiology tests decrease liver biopsy numbers conducted in noncomplex settings where one insult is imagined. For example, various radiographical approaches^10–14 might be adequate for serial monitoring of recurrent or de novo nonalcoholic fatty liver disease in an otherwise stable liver allograft recipient with negative DSA and autoimmune and viral hepatitis serologies. Patients with complex simultaneous insults (eg, suboptimal biliary tract drainage coexistent and TCMR with central perivenulitis) will likely be missed using radiographic monitoring alone and treatment (eg, stent placement versus augmented IS) greatly differs.

To maximize the value proposition of traditional histopathology, pathologists need to comprehensively communicate results and diagnoses for incorporation into patient management algorithms^51,55 (see Appendix S1). This approach is superior to issuance of “descriptive” reports. Granular templated, Banff-based, semiquantitative scoring of microscopic findings^51,55 (eg, portal inflammation and portal, subsinusoidal, or perivenular fibrosis scored: none: 0; minimal: 1; mild: 2; moderate: 3; and severe) standardizes EMR databases (Figure 1). Templated scoring facilitates translational research and enables “digital” queries of archival material, which can be melded with other data (eg, demographics, serology, mRNA expression arrays, urine proteomics) and subjected to analyses.^56–61 Scoring standardization via digital WSI archives facilitates proficiency training and consensus criteria development.⁶² This is an important step toward NGP implementation when objective software-assisted scoring will replace semiquantitative scoring and provide data for DIA algorithm validation. Subsequent NGP “hybrid” approaches will combine human conceptualization skills and “savant-like” DIA capabilites^3,63 to produce “software-assisted sign-out”.

Figure 1.

Overall workflow and NGP integration with other data sources. Locally- or remotely obtained tissue biopsies can be triaged for traditional and NGP and tissue shared with other “invasive” non-histopathology platforms either before or after FFPE processing. WSI can be created at the core laboratory or electronically transferred from remote sites thus saving repeat traditional histochemical staining and preserving tissue for other NGP assays. NGP-generated data can then be integrated with clinical and serological data, noninvasive imaging and fluid -omics and invasive tissue-based -omics platforms for analysis. Dashed lines indicate iterative processes where feedback from a later stage is incorporated to optimize a result.

Continued Value of Traditional Histopathology

Reasons repeatedly cited by advocates for indication and protocol liver allograft biopsies^64,65 and native liver biopsies in the context of hepatocellular carcinoma,⁶⁶ NASH,^52,67 chronic viral,⁶⁸ and autoimmune hepatitis⁶⁹ include: 1) detecting subclinical pathology^{56,57,65,70–72} (reviewed in references 60, 64, 73, and 74); 2) discerning and analyzing pathological events at a micron resolution scale guiding pathophysiological insights^52,55,67; 3) sampling organ-restricted events (inflammation, swelling, etc); 4) serving as an endpoint in management algorithms and clinical trials^49–53; 5) providing detailed meaningful results within hours; and 6) providing tissue samples for other platforms.^{56,57,65,68,70–72} Despite protestations that histopathology is “non-mechanistic,” it discovers, documents, and provides initial pathophysiological insights into many disease processes.^{28,36–38,41,56–58,75–78}

Recently recognized examples where liver allograft biopsy’s value helped advance the field includes recognition and monitoring of: 1) acute^58,59,77,79 and indolent chronic antibody-mediated injury^57,80–83; 2) processes facilitating IS minimization^{49,60,61,84,85} (similar to those used to monitor AIH)^69,86; and 3) regenerative medicine attempts to create and/or rehabilitate donor organs (reviewed in reference 87). Liver allograft recipient management complacency based on relatively normal liver injury tests in long-term survivors ignored decades-long concerning histopathological data of subclinical inflammation and indolently progressive fibrosis (reviewed in references 55, 73, 74, and 82-84). Serum DSA monitoring combined with traditional histopathological analysis showed a robust association among inflammation, fibrosis, and DSA positivity.^56,57,83 Uninflamed biopsies more often are indicative of successful IS minimization.^{49,55,61,72,84,85}

Introduction to Next Generation Pathology

Alternative monitoring approaches should be measured against NGP, which can be broadly defined as the use of multiplex tissue labeling techniques coupled with digital imaging and viewing, automated DIA, and creation and utilization of archival systems (reviewed in references 88-92). NGP enhances the value equation for “histopathology” and can be aggregated with data from other monitoring platforms (Figure 1). Tissue staining can range from standard histochemical stains (eg, H&E, trichrome) to mIHC labeling of multiple (usually two to six) proteins, mRNA, and/or DNA analytes/targets. A “panel” refers to a complete set of stains on one slide, which can be analyzed for spatial relationships among analytes.

NGP increases the value equation for biopsy evaluation by: 1) replacing semiquantitative scoring with objective quantitative data extraction; 2) creating digital archives to feed ML platforms (see below); and 3) leveraging hypothesis-driven translational research results to develop robust, clinically-applicable panels for clinical utilization by a preexisting network of expert pathologists who are familiar with automated immunostaining, standardization, and quality control platforms (eg, CLIA, CAP).

NGP follows the general workflow (Figure 2, Table S1): 1) hypotheses formulation and antigen/primary antibody selection; 2) tissue preparation; 3) slide staining; 4) digital slide scanning; 5) image/stain quality assurance; 6) DIA and biostatistical summary; and 7) digital archiving.

Figure 2.

Next generation pathology workflow and initial panel development. (A) NGP requires hypothesis generation, potential antigenic targets, available antibodies, and preferred detection systems. Tissue preservation and preparation should facilitate staining. Analytics occur simultaneously and iteratively with panel development, scanning optimization, and QA; each step providing feedback to others. Samples are then batch processed using the same protocol/metrics for each sample without human intervention culminating in objective data reporting and archiving. (B) The CD34/PTPRC/HLA-DPB1 multiplex IHC panel: CD34 (green) labels elongated endothelial cells lining the portal microvasculature. PTPRC^high (cyan) mostly labels lymphocytes²⁴⁵ resulting in circular, cytoplasmic “donut”. HLA-DPB1 (red) can appear on any antigen presenting cell (ie, oddly shaped macrophages, round PTPRC positive B cells, and dendritic and endothelial cells). The larger image shows an overlay of all labeled antigens with nuclei shown in blue. See Table S1 for sample workflow details. (C) Avidin-biotin amplification systems are used for sequential detection of proteins via primary antibodies from the same host species (eg, detection of CD34). (D) Tyramide signal amplification systems are used for sequential detection of primary antibodies from the same or different host species. The catalyst horseradish peroxidase activates tyramide-conjugated fluorophores causing them to covalently bind to protein tyrosine residues near the target antigen. Covalent binding enables heat deactivation, denaturation, and elution of primary and secondary antibodies leaving the fluorophore covalently bound to the target antigen and facilitating subsequent staining steps.

Biopsy Procurement and Tissue Preservation

Biopsy value is only as good as the tissue obtained; best practices will ensure that biopsy value can be maximized. At most institutions, interventional radiologists conduct biopsies under ultrasound guidance using biopsy guns. After local anesthesia, the needles are advanced to the organ surface then Tru-Cut needles are triggered to automatically obtain a tissue core (recommended 16-gauge) of a specified depth/chamber depth (22 mm).

When our institution drifted from AASLD standard practice guidelines⁴⁷ of two 2.0 cm 16-gauge (16G) needle cores toward one 1.0–1.5 cm 18-gauge (18G) core the rate inadequate 18G core biopsies increased (approaching 30%), which in turn, prompted a formal study. Thinner gauge needles were originally intended to sample liver malignancies where a greater risk of bleeding exists.^93–95 Our prospective randomized study (n: 150), comparing 16G and 18G cores⁹⁶ showed: 1) no difference in pain or complications rates; 2) shorter 16G core lengths (1.7 cm vs 1.9 cm; P <0.05); and 3) more portal tracts in 16G cores (14 vs. 13; p=0.1). Of 18G biopsies, 29% were inadequate compared to 19% of 16G biopsies (p=0.17), despite 16G cores being significantly shorter.⁹⁷ Two recent larger studies (n: 427; n: 194)^98,99 showed significantly higher inadequacy rates with 18G needles, but Iqbal et al⁹⁸ showed higher pain rates with 16G needles. Myers et al (n: 4275; gauge unreported) showed that 0.75% of patients experienced significant complications: 0.51% pain requiring hospital admission, 0.35% bleeding; and deaths (0.14%) occurred exclusively in cancer patients.¹⁰⁰ Similar results were reported in three other studies,^94,101,95 where terminal complications were restricted to cirrhotic and/or cancer patients.^94,101

Needle gauge significantly affects tissue volume. Nicholson et al¹⁰² compared 16G (0.7 mm) versus 18G (0.5 mm) kidney tissue core diameters. For a 2.0 cm-long biopsy, the cylinder volume formula (πr²h) shows that 16G cores yield double the tissue volume. FFPE tissue mRNA expression profiling can require up to three consecutive 20 µm curls¹⁰³ compared to a single 2–3 μm section for panel staining (see Appendix S2). To maximize approaches available for tissue monitoring, 16G cores should be the standard, and two cores are preferable. Quality metrics for all biopsy-obtaining specialists should address the consequences of: 1) inadequate biopsies; 2) institutional review board approval for a second core; and 3) restricted ability to conduct studies on excess tissue.

Tissue preservation CAP guidelines exist to ensure maintenance of target antigenicity for IHC-based clinical assays¹⁰⁴ (eg, HER2,^105,106 ER and PgR¹⁰⁷). Immediate tissue snap freezing is ideal for antigen preservation but is generally outside of routine clinical workflow and improper cryopreservation leads to inferior morphology preservation for DIA (see Appendix S3). Clinical standard FFPE tissue preparation results in less quality variation and preserves morphology for DIA. Since formalin preparations (10% NBF vs. 4% paraformaldehyde) can produce analytical variability on an antigen-specific basis, a standard should be chosen.^108,109 Room temperature fixation in 10–20x tissue volume of 10% NBF for one hour/mm tissue thickness (generally 6 to 48 hours) is standard for most antigens.¹¹⁰ Preservation-sensitive antigens (ex. p-AKT, p-ERK1/2) require immediate fixation; slicing large specimens facilitates proper fixative penetration^111,112 (See Appendix S2). While no official CAP guidance is available regarding specimen fixation for “molecular pathology” assays, guidelines exist for assay validation from a variety of inputs (ie, unfixed vs. fixed vs. FFPE).^113,114

Formulation of Hypotheses and Antigen/Primary Antibody Selection

Hypothesis testing often leads to innumerable marker queries and panel combinations that can be simplified into key processes. Examples include characterization of immune infiltrates (eg, T, B, and NK cells, and monocyte/macrophage panels)^60,115; relationships of cell types to tissue structures (eg, leukocytes to bile ducts or hepatocytes or capillaries via TBX21/GZMB/CK19/CAM5.2 or PTPRC/CD34/HLA-DPB1, respectively)^115–121; spatial relationships among immune infiltrates (virtual immunological synapses⁶⁰); cancer signaling pathways^118,120,121; stereotypic responses to injury/stress (eg, endothelial and myofibroblast proliferation/activation via CD34/ACTA2/MKI67); and tolerogenic vs proinflammatory liver APCs (eg, IRF4/IRF8/LILRB2/HLA-DPB1/ITGAX). Cancer studies primarily focus on the tumor immune microenvironment – a mirror image of transplant immunology^122–126. Since perturbed genes “travel in herds^7,45“ and are network-linked,¹²⁷ hypothesis-driven or -omics-derived “gene lists” can inform more selective, and potentially more enlightening, NGP targets for study because it enables analyses of spatial relationships. New technologies combine spatial resolving capabilities of digital imaging with tissue-targeted mRNA expression analysis.¹²⁸

Primary antibody or ISH probe selection is key to panel development success. Ideally, primary antibody candidates should have a robust publication use record in human tissues, be easily implemented clinically, and show compatible antigen retrieval characteristics with panel coanalytes. Novel antibodies may not meet this standard, but time is wasted trying to facilitate “poor-performing” antibodies are compounded during attempted panel inclusion. In our experience, ~70% of antibodies show robust FFPE tissue staining; the Human Protein Atlas is a generally reliable reference.^129–131 RTU or IVD antibodies should be considered as the starting point of clinically-targeted panel development⁹² (See Table S1).

Multiplex Staining Approaches

Traditional mIHC approaches include the following (or combinations thereof): (1) simultaneous protein detection via primary antibodies from different host species; (2) sequential protein detection via primary antibodies from the same host species with interspersed blocking steps (Figure 2C); (3) sequential detection of proteins via primary antibodies from the same/different host species with antibody denaturation/elution between detection steps (Figure 2D). Targets can be localized with chromogens (eg, 3,3’-diaminobenzidine [DAB], 3-amino-9-ethylcarbazole [AEC], and alkaline phosphatase substrates¹³² and/or fluorescent probes [ie, organic fluorophores, Opal Dyes, and Qdots]).

Antigen retrieval is usually required for FFPE tissues and impacts panel co-analyte selection because of potential incompatibilities (reviewed in references 133 and 134). Amplification systems (eg, biotin-avidin, tyramide, hapten-based, rolling circle, polymer-based, silver^89,135–137) help to detect low expressing antigens or deterioration-sensitive targets. Newer methodologies, like multiplex ion beam imaging,^138,139 can detect ~100 analytes in FFPE tissues with: (1) improved signal-to-noise ratio and assay linearity; (2) elimination of spectral overlap; (3) analyte signal stability; (4) stable analyte registry; and (5) improved resolution compared to other mass-based reporter systems.¹³⁸ CODEX utilizes oligonucleotide-labelled antibodies, iteratively queried by labelled dNTPs, and visualized by standard fluorescent microscopy.¹⁴⁰ CODEX revealed 66 targets in frozen tissue sections and touts lack of antigen degradation during repeated deactivation.¹⁴⁰ Although new techniques circumvent the need for primary antibodies from unique species, we find it helpful (see Appendix S4). Protocols should be validated in a variety of tissue-specific specimens (ie, test stain targeted tissue; see Appendixes S5–S7).

Implementing mIHC fluorescence panels requires adjustments for possible spectral overlap and/or tissue autofluorescence (see Appendixes S5–S7). Complex mIHC panels may require splitting across several thin serial sections and digitally aligning serial sections. Algorithms can align WSI at regional (eg, portal tracts) and cell-based levels.¹⁴¹ The latter requires more consideration, validation, and manual input to override error. Lopez et al achieved an automated registration accuracy of <50 μm¹⁴² advocating thinner (3–4 μm) section use (see Appendix S2).

Our initial mIHC panel development using Qdots^5,20 was based on indirect detection of primary antibodies using biotinylated secondary antibodies and streptavidin-linked Qdots (Figure 2, B and C; reviewed in reference 143). Initial Qdot advantages included: brightness, photostability, and a common excitation wavelength (acquisition speed facilitated by efficient metal halide excitation sources). Qdot shortcomings became apparent over time, including: lot number staining variation; signal quenching by a variety of commonly used IHC reagents; and background staining (especially in liver, see Appendix S7), which led to unreliable staining for some panels.^144,145 The availability of enhanced camera sensitivity led to a transition to fluorophore-conjugated TSA reagents (Figure 2D, for caveats see Appendix S8).^{122,123,146,147} Heat deactivation and elution remove primary and secondary antibodies from the antigen binding site (to below detection levels) at each staining step. Fluorophores remain covalently bound to tyrosine residues near the antigen site via peroxidase activation.^123,136,148

Automation of Multiplex Immunohistochemistry

Automation is essential to mIHC clinical implementation because it reduces human variation and errors while producing more predictable results and shorter turnaround times: week versus day(s) and is available in most clinical laboratories. We have experience with two platforms – LabVision Autostainer 360–2D (ThermoScientific) and Discovery ULTRA (Ventana Medical Systems/Roche). The LabVision is suited to avidin-biotin based detection systems but is limited compared to newer platforms (eg, Discovery ULTRA, Leica Bond III) by inabilities to: 1) deparaffinize; 2) enable individual protocolized slide heating for antigen retrieval or incubations (as needed for TSA and polymer-based detection reagents and ISH), which may shorten primary antibody incubation time; 3) prevent evaporation during lengthy incubations (>2–3 hours); and 4); enable more than one staining protocol per run. Newer platforms enable completion of mIHC four to five analyte panels in “overnight runs (8–15 hours)”. “Open” platforms allow investigators to customize their reagents in conjunction with vendor-specific reagents.

WSI Considerations

Clinical NGP imaging considerations include: 1) overall system costs (ie, money and time)¹⁴⁹; 2) workflow integration; 3) WSI resolution; 4) scan time for transmitted light microscopy applications; 5) fluorescence scanning capabilities; 6) digital storage space; 7) trained personnel; 8) LIMS integration; 9) telepathology; and 10) regulatory hurdles.^150–152 Comprehensive reviews cover selection, implementation, and practical use of digital pathology systems in clinical practice or research settings.^153–155 Industry advocate groups (eg, CAP, Royal College of Pathologists) and software workflow improvements drive standardization and validation guidelines with respect to device and viewing platform calibration, maintenance, analytics, and validation cohort size,^156–158 In 2017, the USFDA granted Philips’ IntelliSite scanner approval for primary diagnostic use.¹⁵⁹ “Subscription-based” companies/services offer economical alternatives to a large upfront capital or personnel investments.^160,161

Retrospective studies confirm diagnostic concordance between WSI and glass slides using the industry clinical standard of 20x WSI magnification (0.5 microns/pixel resolution).^162,163 Higher 40X magnification scanning (<0.25 micron/pixel resolution) is: 1) required for transplant and “inflammatory” pathology (eg, recognition of leukocyte phenotypes and glomerular basement membrane abnormalities)^164,165; 2) supported by nearly all WSI vendors; 3) alleviates pathologist eye strain compared to 20X; and 4) yields over 95% diagnostic concordance compared to conventional light microscopy.¹⁶⁶

Fluorescence WSI processes require upfront mIHC experimental design considerations such as fluorophore/imaging filter pairs with available excitation sources to reduce imaging artifacts such as photobleaching and poor signal-to-noise. Web accessible resources can assist with optimal component selection.¹⁶⁷ Improvements enabling routine use of fluorescence WSI, include: camera technologies,¹⁶⁸ fast excitation sources, wavelength separation,¹⁶⁹ improved axial resolution (Z-dimension) techniques (ie, confocal WSI scanners), and mechanical speed.¹⁷⁰ Emerging high-speed focus systems can leverage new camera technologies to shorten capture times,¹⁷¹ and correlative microscopy approaches will enable comprehensive multi-modal imaging studies of tissue events.¹⁷²

Fluorescence WSI scanners enable comprehensive software-based mIHC quantification (eg, NearCYTE¹⁷³; Figure 3) and social relationships among cell types^60,119,174 (Figures 3 and 4). Graphical mIHC data displays can mimic flow cytometry while preserving spatial context and “tether” index populations back to tissue locations¹¹⁹ (Figure 5). Brown et al demonstrated equivalence between mIHC and traditional flow cytometry for assessing lymphocyte populations.¹⁵⁸ A mIHC WSI enables systematic “reanalysis” of tissues as insights are gained from analytic datasets. The potential for combining fluorescence and brightfield imaging to predict tissue labeling “in silico” will extend spectral sensitivity.¹⁷⁵ Emergence of immune checkpoint inhibitors in cancer therapy will trigger widespread adaptation of digital imaging capabilities.

Figure 3.

An example of objective quantitative analysis of replacing semiquantitative scoring. (A) An H&E image of a biopsy with obvious portal lymphocytic inflammation (top right) would be scored semi-quantitatively as mild to moderate portal inflammation. The lymphocyte phenotypes and quantitative analysis of each cell type are undiscernible from this traditionally stained and scoring approach. However, human only analysis can be improved by mIHC staining and automated DIA. (B) For example, CD3(green)/MAC387(red)/CD163(cyan) panel can be used to detect and quantify total T cells (CD3) and recent monocyte/macrophage immigrants (MAC387) signaling some type of liver injury,^246,247 and resident Kupffer cells (CD163). The red box identifies the high-power field shown in D. (C) A combination of pixel-based and QH can identify individual cells and their phenotypic characteristics based on analyte expression, which can be used to objectively report the total number of each cell type per mm² of biopsy area or in certain tissue areas (eg, portal tracts). Here we localize nuclear (blue) or cytoplasmic (green) analytes in possible identified cytoplasmic regions (grey) per segmented nucleus. (D) The number of cells matching the criteria defined in C are then indicated by Nearcyte, a hand-crafted” tissue-tethered cytometric software and can be quantified. (E) An unsupervised clustering algorithm can be used to identify and highlight abnormal “clustering” of CD3⁺ T cells using objective computational algorithms, such as mixed gaussian models (yellow ovals). The proximity to and type of APC present in these clusters, like MAC387⁺ or CD163⁺ cells, can be analyzed to determine whether recent immigrants or liver-based, possibly more tolerogenic, cells, respectively, are interacting with tissue-based CD3⁺ cells. MAC387⁺ cells are indicated with green dots.

Figure 4.

Digital image analysis. (A) and (B) are from two patients at time 0 (before attempted IS weaning). Overall the severity of the negligible portal and lobular inflammation is similar based on analysis of the H&E-stained sections. However, when the same biopsies (C) and (D), respectively, are mIHC stained for CD34 (green), PTPRC (cyan), HLA-DPB1 (red), and QNuclear Deep Red stain (blue) the stark difference between the two is easily recognizable. (C) DIA to identify pairings (virtual immunological synapses) between APC and PTPRC+ cells (yellow circles) highlights a vast difference between the two biopsies that is not recognizable on the H&E stains. Note the noticeable difference between (C) and (D) that is not noticeable on the H&E stains. (E) DIA is also able to recognize portal tract regions defined as a 100 μm radius zone centered on 5 or more HLA-DPB1+ (red) cells within a 25 μm area (large thin yellow circles). (F) DIA is able to identify “virtual immunological synapses” formed between two cells that may not be apparent to the human eye, in this case a CD34⁻/PTPRC⁻/HLA-DPB1⁺ (red) cell within 5 μm of a CD34⁻/PTPRC⁺/HLA-DPB1⁺ (leukocyte) cell (small thick yellow circles). Quantitation of these features can be queried and combined with spatial information to infer relationships about antigen presenting cells in portal tracts.

Figure 5.

Multidimensional morphometric data generation and tissue-tethered cytometry. DIA can dissect tissue cell populations and produce data outputs mimicking flow cytometry, while maintaining the spatial context of mIHC panels. Automated detection of tissue area, nuclear segmentation, and analyte expression/cell are performed via NearCYTE. This mIHC panel included CD4 (green), CD8 (red), FOXP3 (magenta), TBX21 (cyan), and nuclei (blue) to help characterize different T cell phenotypes. (A–F) Cellular subpopulations can be graphically represented on the mIHC-stained tissue sections: (A; CD4⁺/FOXP3⁺ cells; yellow circles) and (B; CD4⁺/TBX21⁺ cells; red circles). (C) Overlay of all mIHC panel components with DIA can be used to quantify cell types and inform relationships among cell types and tissue structures. Morphometric data, maintaining spatial relationships between cell types, can be translated into a traditional flow plot with the color intensity representative of the percentage of cells (green less cell density to red more cell density). (D) Cell plot representing CD4⁺ versus CD8⁺ cells from all cells that had at least one mIHC panel label. (E) CD8⁺ cells were queried for subpopulations expressing TBX21. (F) CD4⁺ cells were queried for subpopulations into putative Tregs expressing FOXP3 or Th1 cells expressing TBX21.

WSI quality assurance for DIA is a time-consuming, but essential, two-component process needed to eliminate: 1) scanning errors (eg, focus problems, tissue folds, etc)^176,177 and 2) inadequate staining. Scanning errors are being minimized by algorithmically incorporating color and focusing “metrics” to automatically detect problematic areas.^178,179 Staining quality assessment requires domain knowledge of expected staining patterns, which has been more difficult to automate.

Digital Imaging in Transplantation and Analytical Approaches

Digital imaging and DIA have a 25-year history in clinical translational transplantation research and enables a worldwide network of expert pathologists. Digital images have been used to evaluate the reproducibility of diagnoses and/or lesion scoring in kidney^180–182 and heart¹⁸³ allografts and C4d deposition in liver¹⁸⁴ and kidney¹⁸³ allografts. DIA approaches included: 1) pixel-based; 2) QH; 3) combinations of pixel-based, spatial and QH (ie, hand-crafted); and 4) ML. When compared with estimation techniques (eg, stereology), WSI-based analyses have provided results with acceptable correlation to human-generated assessment.^185,186

Pixel-based “software recognition” of specified color(s) differences has been used to quantify: 1) fibrosis in liver,^61,187–191 kidney,^192–196 and heart^173,197 allografts with stains that enhance fibrous and non-fibrous tissue contrast (eg, Masson’s and light-green trichrome, Picrosirius Red with/out polarization, collagen III immunostaining); 2) severity and phenotype of inflammatory infiltrates in kidney^{115,198–200} and heart²⁰¹ allografts; 3) severity of steatosis in donor livers^202–205; and 4) scoring of chronic allograft damage index²⁰⁶ parameters. Results show that human semiquantitative scoring correlates with software-based DIA, but the latter is generally more reproducible - a key criticism of human-only traditional evaluations. Donor liver^202–205 steatosis severity evaluation, augmented by Oil Red O lipid staining²⁰⁵, might benefit from DIA standardization because humans tend to overestimate the severity of steatosis compared to software-derived values.²⁰³ Since pixel-based analysis is not computationally taxing, readily available devices such as smartphone cameras can be leveraged to assess semi-supervised steatosis quantification in donor liver biopsies.²⁰⁷ Once digital fibrosis quantification tools are standardized, they will erode the subjective and poorly reproducible criticisms of histopathology.

Although pixel-based approaches are useful and simple, human intervention can be required, generated data can be limited, and spatial contexts are generally ignored. For example, pixel-based assessment of liver^187–190 and kidney^{192,194,195,206} allograft fibrosis and kidney^{115,198–200} and heart²⁰¹ inflammatory infiltrates often rely on humans either capturing and/or excluding ROI (eg, glomeruli, blood vessels, etc) or segmenting/selecting cortical tissue on WSI, which introduces subjectivity. Biopsies from architecturally normal or minimally fibrotic livers can yield a wide range of quantitative fibrosis expression because of sampling of variably-sized portal tracts; therefore, an assessment of architectural integrity is needed.^191,208,209

QH groups similarly colored pixels into biologically relevant structures (eg, nuclei) that satisfy relevant predefined criteria (eg, shape and/or size). Preserved spatial data (eg, XY coordinates, morphological characteristics of discrete structures) requires high sensitivity and specificity detection algorithms, limiting adaptability due to routine error sources (ie, tissue artifacts or stain variation²¹⁰). QH has been used for glomerular parameterization,^211,212 fibrosis assessment²¹³ and/or determining total infiltrative cell types and/or characteristics of particular cell types (Figures 5 and 6) within tissues.^{60,115,118,119,121,214,215} Quantitative DIA evaluation of nephropathology protocol biopsies for interstitial fibrosis and tubular atrophy lesions has shown prognostic significance.²¹⁶ Assessment of assay reproducibility is essential for standardization and determining usefulness.

Figure 6.

Vendor-neutral web-based viewing solution. This vendor-neutral viewing platform enables direct side-by-side comparisons of biopsies from the same patient at two different time points. The “Weaning/Enrollment Biopsy” (left) and “Biopsy at 16 months post-weaning (right) show representative portal tracts from H&E-stained (top row) and CD34/HLA-DPB1⁺/PTPRC^high stained mIHC panel (bottom row). Quantitative queries, analyses, and mechanistic insights can be carried out with respect to treatment changes between the two timepoints. This platform also enables receipt of WSI created elsewhere thus eliminating the need to repeat common stains (H&E, PAS, trichrome, etc) preserving tissue for other analyses.

Complex DIA combines pixel-based and QH analyses into higher order relationships using rule sets to define tissue structures (eg, algorithmic detection of “virtual immunological synapses” in pediatric liver allografts distinguished among patient clusters enrolled in IS minimization trials; Figures 3 and 4).⁶⁰ Madabhushi et al discuss the concept of “hand-crafted” DIA features, or “features which can be connected to specific measurable image attributes and have some degree of interpretability.”²¹⁷ They report interest in “graph-based” approaches (ie, using tessellations and other human interpretable spatial relationship metrics) to characterize image feature arrangement (eg, arrangement of nuclei).^217,218 Recognition of specific pathology patterns (eg, lymphocyte clustering; Figure 3) can correlate with downstream effects (eg, rejection).²¹⁹ Standardization of thresholding, deconvolution, and automated ROI selection can reduce variability in quantification compared to “human-powered” reproducibility.^220,221

We^{20,60,61,116,118–121,214} and others^{90,122,222–225} utilized a combination of pixel-based and QH based techniques to analyze: 1) graft-infiltrative cells and relationships between and among them (Figure 5); 2) tumor immune microenvironments; and 3) side-by-side comparison of biopsies from the same recipient over time (Figure 6). Automated detection of tissue area, nuclear segmentation, and analyte expression/cell are performed via NearCYTE, a “hand-crafted” tissue-tethered cytometric software (Figure 5).¹⁷³ Multidimensional individual cell-based morphometric data (eg, size, shape, analyte expression pattern, XY coordinates, etc) can identify specific cell populations (Figure 5A-C), and subpopulations (Figure 5B-E) while maintaining spatial relationship with other cell types (ie, tissue-tethered cytometry¹¹⁹). Specific cell populations and interactions can be interrogated within a spatial context while retaining architectural features (eg, virtual immune synapses in portal tracts⁶⁰). This approach can run “asynchronous” to pathology review and provide supporting diagnostic material (eg, density specific cell populations per mm² of area).

DIA creates a wealth of digital data including XY coordinates, color intensity, and size, and can be used to determine spatial relationships among cells (eg, Treg, NKT cells) and/or structures (eg, portal tracts, glomeruli) not readily appreciated and/or quantifiable by humans. Humans are “handicapped” when attempting to utilize multidimensional “big” data resources, because the data inputs are not readily amenable to routine representation methods for pattern detection. A subset of Artificial Intelligence: Machine Learning (ML) enables algorithms to “learn” by observation of input variables that are tied to particular outcomes. ML can then “predict” outcomes on fresh data sets,^226,227 which has the potential to eliminate human imposed constraints and biases.

ML methods are classified by algorithmic “learning” approaches to input data: supervised and unsupervised.^226,227 Supervised learning, the most common, “learns” by algorithmic mapping of training data variables to a known outcome (eg, variables that create a “formula” to give a known output).^226,227 A practical biopsy-centric application of supervised ML includes Content Based Histopathological Image Retrieval, where WSI archives are searched for image patterns (via color, texture or other measured attributes) of a chosen pathology (ie, interface hepatitis lesions or FSGS lesions in kidneys).^228,229 Modern ML can “blend” both supervised and unsupervised techniques to optimize performance.^226,227

“Deep Learning” (DL; deep convolutional network) is a supervised ML method^226,227 that is employed successfully in the healthcare arena (eg, the IBM Watson Oncology tool; https://www.ibm.com/us-en/marketplace/ibm-watson-for-oncology). DL creates data information relationships at multiple levels, similar to humans (reviewed in reference 230), can incorporate absolute data values, and combine them with “abstractions” such as how quickly those values are changing. Benefits include higher training efficiency and leveraging of cloud-based parallel computing resources for performance (eg, nuclear classification of colon cancers to predict outcome).^231,232

ML techniques utilizing unsupervised learning do not rely on any predetermined labeling of training or outcome data, but instead depend on assumptions to discover patterns.^226,227 Unsupervised learning-based ML challenges include: 1) methods that are typically computationally expensive; 2) subtle pathology patterns (eg, “subclinical” rejection) are not well characterized making training and validation challenging, especially considering biopsy size and architectural variance; and 3) influence of data artifacts, such as histology or intra-laboratory specimen preparation variance.

Unsupervised approaches operate bias-free, and hold promise to unlock biological/pathological signatures (Figure 3E). Engagement and inclusion of pathologist-oriented insights in unsupervised learning processes provide unmatched value to minimize caveats²³³: 1) reducing non-specificity from “savant” DIA data algorithms (eg, training data selection to avoid histopathology artifact contribution), and 2) predictions grounded in observable and understandable facts.²³⁴

Regardless of the ML approach, pathology practices will inevitably evolve toward ML incorporation into the routine workflow to maximize specimen value. Practical applications will enable “probing” biological questions by incorporating causal modeling, distinguishing which variables influence the prediction process. Routine screening tasks (eg, fibrosis-based architectural distortion quantification) will be automated, promoting efficiency and decreasing prognostic variations for subtle changes. Pathology classifications will evolve to incorporate these subtleties, invisible to the human eye, to resolve diagnostic dilemmas (eg, invasive versus in situ carcinomas).

Digital Archiving

WSI scanners create “pixel” characteristic data from stained slides, leading to a requirement for scalable and cost-effective data storage (eg, our digital archives occupy 80+ TB). WSI data contains or is linked to, metadata (eg, semi-quantitatively scored parameters, demographics, serological laboratory data, -omics platform data, etc), including associative details undiscernible to humans, but informative to ML (Figure 1). Linking metadata and intra-WSI pixel data is critical to full NGP utilization and integration into the clinical workflow. Archival systems should be user friendly and support multiple data sources. Commercial scanner vendors offer image management systems, but they: 1) are often rigid and relatively closed, making it difficult to integrate with existing laboratory information systems; and 2) have considerable upfront and ongoing maintenance costs.

A WSI archive of longitudinally-obtained biopsies and associated metadata yields several benefits because it enables: 1) immediate availability of prior patient biopsies regardless of pathologist location; 2) side-by-side biopsy comparisons, special stains, and fluorescence microscopy (Figure 6); 3) lesion grading accuracy by comparing to consensus archival reference standards; 4) remote access for expert consultation; 5) immediate harvesting of electronically-defined groupings for mechanistic, translational research, and quality assurance initiatives; and 6) batch processing DIA comparisons between or among groups to test various hypotheses and for ML algorithms. For example, simply by changing selection criteria for group inclusion (eg, DSA+ versus DSA-; male versus female, etc) various WSI parameters (eg, portal inflammation) can be batch processed for comparison without having to rearrange and link associated metadata – even for parameters not previously evaluated.

Industry advocate groups (eg, Digital Imaging and Communications in Medicine) promote image storage formats and data exchange²³⁵ standards that facilitate WSI exchange and interoperability.²³⁶ Open source WSI software, image format interpretation, data collection and analysis frameworks enable high throughput algorithm execution and downstream organization.²³⁷ The Open Microscopy Environment²³⁸ and Digital Slide Archive²³⁹ groups developed storage models and database systems that can represent and manage multi-dimensional WSI data. Cloud-based computing platforms are being leveraged for solutions that integrate search, storage, and DIA processing in a secure environment that can be cost-effectively deployed.^240,241

NGP Translation Into a Clinical Transplant Pathology Workflow

Clinical translation of NGP platforms is needed to exploit its capabilities by distributing capabilities amongst a preexisting worldwide network of expert pathologists. Our core laboratory for multicenter and corporate-sponsored clinical trials transitioned more than a decade ago to an entirely digital workflow, including: complete barcode accessioning and tissue and slide tracking, WSI creation, automated WSI analyses, and archiving. This prototypical clinical setup has been successful primarily because of: 1) the relatively small daily case load (about 20 biopsy slides/day including standard and mIHC stains); 2) the relatively small size for the vast majority of specimens (liver, kidney, and heart biopsies), which translates into a “controllable” scanning real estate; 3) parallel Zeiss Axio Scan.Z1 WSI scanners; and 4) a team of domain experts devoted to each aspect of the NGP workflow.

Histochemically-stained biopsies are scanned quickly (~11 mm2 at 7 min/WSI at 40x, 0.11 μm/pixel resolution) whereas a four-antigen mIHC WSI biopsy might require 20–30 min/WSI. Larger specimens (eg, sections from failed allografts) require a lengthy scan time (179.74 mm2 at 66 min/WSI). In our experience, mIHC IF scanning may be reliably accomplished with a 20x apochromatic objective; yielding optimal balance between scan time and image quality (signal-to-noise), with a wider field providing greater context for focus determination. The value equation of creating WSI for all specimens in a routine clinical workflow is questionable at this time. All WSIs are subjected to intense, time consuming scrutiny for semiquantitative scoring, which will be replaced by objective “analytical modules” in the viewing software.

Problematic areas for clinical implementation include: 1) lack of a best-of-breed technology to address all considerations for an optimal WSI system and workflow; 2) a paucity of competent trained personnel; 3) LIMS integration issues²⁴², and 4) vendor-supplied WSI viewer incompatibility. Our development of a vendor-neutral web-based viewer complete with annotation and analytical capabilities addressed the last issue and eliminated user interface inefficiencies (Figure 6). Template development with WSI libraries specific to pathology scoring and studies will assist in LIMS system interface and improve standardization and expertise distribution.

NGP penetration into routine clinical patient care has been limited and requires commitment from anatomic pathology groups and clinical staff they support (eg, surgeons, hepatologists, nephrologists). Transplant biopsies require STAT results with “hours” turnaround time for diagnostic rendering. The necessary infrastructure for obtaining and properly triaging needle biopsies already exists and includes core(s) triaged for FFPE (1.5 – 2.0 cm needle cores) and a smaller needle core segment (3–5 mm) triaged for frozen section C4d staining. Although actual pathologist slide review time via WSI is slightly increased compared to glass slides for routine biopsy stains (personal observations and references 180 and 182), the time cost of implementing digital imaging for routine semiquantitative evaluation would currently not exceed two hours. Overnight mIHC staining and quantitative DIA would necessitate that NGP-derived data would require one to two days for more detailed objective data output (eg, pathogenic mechanisms, detailed assessments of the microvasculature.¹⁷³ fibrosis, other architectural distortions).

Developmental arms of clinical immunostaining laboratories will be responsible for panel development, similar to the current workflow. Once validated, panel staining automation and QA by trained laboratory personnel can be implemented with overnight turnaround times for as many as four to five protein analyte panels. Staining platform vendors currently encourage chromogenic mIHC panels for clinical application because of limited imaging and DIA platform availability in clinical labs and pathologist familiarity with chromogenic labeling. Advantages of fluorescence labeling include: 1) precise signal quantification; 2) increased sensitivity; 3) spectral channel separation; 4) standardization; 5) increased number of color options, and 6) facilitation of DIA.²⁴³ Continued cost reductions in equipment and reagents will entice clinical labs to upgrade imaging and analytical capabilities. Care must be taken to ensure that mIHC panels are applied in a medically necessary way and subjected to standards and proficiency testing to be cognizant of patient costs.²⁴⁴

Conclusions

Our view of platforms “competing” with traditional and NGP histopathology is a welcoming one, especially during discovery and developmental phases because radiological, and -omics platforms are complementary. Computer hardware and software advances are driving a revolution in histopathology, which we have overviewed and illustrated its value. In our opinion, although fewer biopsies will likely be obtained, biopsies will be subjected to more intense scrutiny using NGP, which: 1) preserves spatial and inferred temporal contexts; 2) can rely on a worldwide distributed network of expert transplant pathophysiologists who, instead of relying on semiquantitative data rendering on biopsies, will have new toolkits that standardize objective quantitative output that limits or eliminates subjectivity and enhances reproducibility; and 3) engage various non-medical experts (eg, engineers, mathematicians, statisticians, and other biologists) into the realm of tissue analysis.

Integration of multiple data sources filtered through a variety of domain experts will provide a conceptual framework for scientific and patient care advancements. “Iterative loops” comparing morphometric, clinical, and -omics platforms data are informative and creative. Reliance on human input for morphology will continue, but software tools and ML will contribute to final biopsy interpretations once large WSI libraries are created and queried. Retention of spatial and inferred temporal context is essential because: 1) humans use visual inputs to conceptualize their environment; 2) it engages a worldwide network of expert mechanistic pathophysiologists who can interpret and develop in parallel -accelerating the rate of advancement; and 3) it retains software-based DIA and ML biopsy analyses. Until “grind and bind” or radiographic approaches discover methods to reconstruct the original spatial configuration from expression data or achieve the submicron spatial context resolution, respectively, both traditional histopathology and NGP are needed and will continue to be valued.

ABBREVIATIONS:

AASLD

American Association for the Study of Liver Diseases

AMR

antibody-mediated rejection

APCs

antigen presenting cells

CAP

College of American Pathologists

CLIA

Clinical Laboratory Improvement Amendments

DIA

digital image analysis

CODEX

Co-detection by InDEXing

DSA

donor specific antigen

EMR

electronic medical record

FFPE

formalin-fixed, paraffin-embedded

H&E

hematoxylin and eosin

IHC

immunohistochemistry

IS

immunosuppression

ISH

in situ hybridization

IPEX

chromogenic immunohistochemistry

IVD

in vitro diagnostic

mIHC

multiplex immunohistochemistry

ML

machine learning

LIMS

laboratory information management system

mRNA

messenger ribonucleic acid

NASH

nonalcoholic steatohepatitis

NBF

neutral buffered formalin

NGP

next generation pathology

NK cells

natural killer cells

PAS

periodic acid-Schiff

Qdots

quantum dots

QH

quantitative histomorphometry

RTU

ready to use

ROI

region of interest

TCMR

T cell-mediated rejection

TSA

tyramide signal amplification

TTC

tissue-tethered cytometry

USFDA

United States Food and Drug Administration

WSI

whole slide images