A Tissue Systems Pathology Test Detects Abnormalities Associated with Prevalent High-Grade Dysplasia and Esophageal Cancer in Barrett's Esophagus

Barrett's esophagus is a precursor to esophageal adenocarcinoma (EAC), which is the fastest growing cancer type by incidence in the United States with 5-year survival rates of 18% (1). EAC can be prevented if dysplasia is detected and treated early with endoscopic therapies such as radiofrequency ablation (RFA) and/or endoscopic mucosal resection (EMR; refs. 2–4). Current guidelines from the American College of Gastroenterology recommend surveillance by endoscopy with biopsies at intervals determined by the pathologic diagnosis (5). The diagnosis of dysplasia in Barrett's esophagus is limited by the random nature of endoscopic sampling, which may miss dysplastic areas, and by interobserver variation (6). Although subtle lesions containing high-grade dysplasia (HGD) and EAC can be detected by expert endoscopists at high-volume centers, recognition of subtle lesions is more challenging in the community setting (7). These limitations can result in repeat endoscopies and delayed diagnoses of HGD and EAC (8).

A field effect has been described in many different cancer types, including in EAC (9, 10). Dysplasia and EAC can be multifocal in Barrett's esophagus. The same mutations and aberrant DNA methylation have been found at multiple levels and in large fields in Barrett's esophagus (11, 12), indicating field cancerization. A preneoplastic field surrounding HGD or EAC may appear histologically nondysplastic (ND) or low-grade dysplasia (LGD) but exhibit molecular and cellular changes associated with malignant transformation. Detection of abnormalities in this expanded field may overcome the limitations of random sampling and subjective diagnoses, enabling earlier diagnosis of HGD and EAC.

Many biomarkers have been evaluated in Barrett's esophagus (13–16) and the British Society of Gastroenterology (BSG) recommends p53 IHC to aid diagnosis of dysplasia (17). However, no biomarkers have been validated to reliably detect the field effect or abnormalities associated with prevalent dysplasia and EAC in Barrett's esophagus. A tissue systems pathology approach based upon an imaging platform that quantifies both epithelial and stromal abnormalities has been shown to aid in distinguishing HGD from nondysplastic Barrett's esophagus with reactive atypia (18, 19). This imaging approach has also been demonstrated to predict incident progression in Barrett's esophagus, by objectively quantifying molecular and cellular features that precede definitive morphologic changes (Fig. 1; ref. 20). The assay employs multiplexed immunofluorescence labeling of 9 epithelial and stromal biomarkers in sections from formalin-fixed paraffin-embedded (FFPE) biopsies. The fluorescently labeled slides are imaged by whole-slide fluorescence scanning, and automated image analysis software extracts quantitative expression and localization data on the biomarkers and morphology. The final step uses a multivariable classifier to integrate the quantitative image analysis data into individualized scores that are correlated with risk of HGD/EAC (Fig. 1; ref. 20). This may have applications in detecting molecular and cellular changes in the expanded preneoplastic field associated with HGD/EAC. The aim of this study was to determine whether this assay can detect abnormalities indicative of a field effect in ND, indefinite for dysplasia (IND), and LGD biopsies from Barrett's esophagus patients with prevalent HGD/EAC.

A case–control study was constructed that utilized a multicenter cohort of Barrett's esophagus patients with clinical outcome data from four high volume institutions [Geisinger Health System, University of Pittsburgh, University of Pennsylvania, Philadelphia, PA and Academic Medical Center (AMC), Amsterdam, the Netherlands]. Barrett's esophagus cases with ND, IND, or LGD confirmed by gastrointestinal subspecialist pathologists (J.M. Davison, J. Li, and N.C. Jhala) were retrieved. For patients with multiple biopsy levels taken at the same endoscopy, the biopsy with the highest diagnosis determined by a gastrointestinal subspecialist pathologist was selected (LGD was the highest diagnosis, then IND, and ND was the lowest). For patients with multiple biopsy levels with the same diagnosis, the pathologist at each institution selected a representative block with sufficient tissue for analysis. Inclusion criteria were availability of tissue blocks and clinicopathologic data, and confirmation of intestinal metaplasia by a gastrointestinal subspecialist. Exclusion criteria were insufficient tissue quality (assessed by a pathologist), and use of Bouin fixative or methylene blue in sample processing that can interfere with fluorescence immunolabeling. Cases were patients who had HGD/EAC on repeat endoscopy in <1 year (n = 23) or had prior history of treated HGD/EAC, returned to ND, IND, or LGD and had HGD/EAC on repeat endoscopy (n = 7; prevalent cases, n = 30 in total). Prevalent cases with and without a prior history as described above were included as both subsets of patients can harbor HGD or early EAC that can be challenging to recognize during endoscopy. The nonprogressor controls did not show HGD/EAC on repeat endoscopy and had median HGD/EAC–free surveillance time of 5.6 years (n = 145). Data elements collected were: case collection date, original pathologic diagnosis, and GI subspecialist diagnosis for the case tested in this study, date and original diagnosis of every surveillance biopsy, progression endpoint (HGD/EAC), HGD/EAC–free surveillance time (time between case tested and HGD/EAC diagnosis or last follow-up), age, sex, and segment length (cm) and segment class (short ≤3 cm, long >3 cm). The study was approved by the institutional review boards at each institution.

Five-micron sections of FFPE Barrett's esophagus biopsies were stained with hematoxylin and eosin (H&E) by standard histology methods. K20, p16INK4a, AMACR, p53, HER2/neu, CD68, COX-2, HIF-1α, and CD45RO were labeled by multiplexed immunofluorescence according to previously described methods (19). The biomarkers were multiplexed in subpanels of 3 primary antibodies per slide detected via Alexa Fluor-488, -555, and -647–conjugated secondary antibodies and Hoechst-33342 to label DNA (Life Technologies).

H&E-stained slides were imaged at ×20 magnification on a NanoZoomer Digital Pathology scanner (Hamamatsu Photonics). Fluorescently immunolabeled slides were imaged using a standard operating procedure at ×20 magnification on a ScanScope FL (Leica BioSystems) as described previously (19).

Whole-slide fluorescence images were analyzed using the TissueCypher Image Analysis Platform (Cernostics, Inc.), which utilizes automated tissue image analysis algorithms for segmenting cell-based objects and tissue structures (e.g., epithelial and stromal compartments) to allow contextual, quantitative biomarker, and morphology feature data collection. The image analysis algorithms have been described in detail previously (19) and are summarized in Supplementary Fig. S1. The 15 features employed by the risk classifier (Supplementary Table S1; ref. 20) were extracted from the fluorescence whole-slide tissue images.

A risk prediction classifier was developed in a previous study for prediction of incident progression to HGD/EAC (Fig. 1; ref. 20). In this study, we tested the hypothesis that the patients in the predicted high-risk class have significantly higher risk for presence of prevalent HGD/EAC than patients in the predicted low-risk class. We also tested the hypothesis that the risk classes would provide independent and stronger prognostic information beyond that of the pathologic diagnosis (GI subspecialist or generalist pathologist). Sample size calculations indicated that a total of 43 patients (including both prevalent cases and non-progressors) were required to ensure 80% power to detect a significant difference of 50% in the risk of prevalent HGD/EAC between those classified as high-risk vs. low-risk, at a 0.05 significance level. All assay parameters were prespecified, including the 15 image analysis feature/measures, scaling parameters, the classifier model, and cutoffs as defined in the previous study (20). The assay parameters are summarized in Fig. 1 and Supplementary Table S1). The risk score and risk class (low, intermediate, or high) were calculated for each case.

ROC curves were plotted on the basis of the binary outcome of the subsequent diagnosis of HGD/EAC (cases) versus no disease progression (control) and the continuous risk scores of the test. ROC curves were also plotted for percentage of cells overexpressing p53 (determined by the image analysis software as described previously; ref. 19). The comparison to p53 was done because the BSG recommends p53 IHC to aid in the diagnosis of dysplasia (17). Logistic regression was used to evaluate the significance of association of the predicted risk groups as the independent variable with subsequent diagnosis of prevalent HGD/EAC or not as the dependent variable. ORs with 95% confidence interval (CI) measuring the strength of the association between the predicted risk groups and the subsequent diagnosis of HGD/EAC were calculated from the logistic regression.

Multivariate logistic regression and multivariate Cox regression were performed to compare the performance of the risk classes produced by the classifier versus the pathologic diagnosis by either generalist or GI subspecialist included as the independent variable, in predicting subsequent diagnosis of HGD/EAC included as the dependent variable. The pathologic diagnosis was dichotomized: LGD versus ND and IND combined. IND and ND cases were combined due to the limited sample size of the IND subset (2 non-progressors and 1 prevalent case had subspecialist diagnosis of IND).

The case–control cohort included biopsies with diagnoses of ND (n = 13), IND (n = 1), or LGD (n = 16) from 30 Barrett's esophagus patients with HGD or EAC (prevalent cases, median time to HGD/EAC diagnosis 140.5 days; IQR, 56–241) and 145 samples from matched control patients with clinical outcome data showing no disease progression (ND n = 138, IND n = 2, LGD n = 5, median HGD/EAC-free surveillance time 2,015 days, IQR 1,498, 3,111). 22/30 prevalent cases were diagnosed with HGD and 8 of 30 were diagnosed with EAC on repeat endoscopy (Table 1). The control patients were from a cohort evaluated in a previous study (20), whereas the prevalent cases had not previously been evaluated. The majority of the patients were male, and a higher proportion of patients with HGD/EAC were male (93.3%) and had long segment Barrett's esophagus (63.3%) compared with the non-progressors (78.6% were male, 50.3% had long segment), which is consistent with published epidemiology studies in EAC (21, 22). The clinical characteristics of the patients are summarized in Table 1. 

The prespecified 15-feature risk classifier was evaluated in the set of Barrett's esophagus biopsies from prevalent cases and non-progressor patients. ROC analysis based on the binary outcome (subsequent diagnosis of HGD/EAC vs. no disease progression) and the continuous risk scores showed that the classifier had the capability to distinguish prevalent HGD/EAC from non-progressors with AUROC of 0.893, whereas the percentage of cells overexpressing p53 had AUROC 0.594 (Fig. 2A). Subanalyses were performed on the subsets of prevalent cases with and without a prior history of HGD or EAC, as cases with a prior history of HGD/EAC may harbor greater numbers of mutations and other abnormalities than cases with no prior history. ROC analysis of the subsets of prevalent cases with and without a prior history of HGD or EAC (n = 7 and 23, respectively) showed that the classifier had strong predictive performance in distinguishing both subsets from non-progressors (AUROC = 0.926 and 0.883, respectively). AUROC for the ND/IND and LGD subclasses were 0.873 and 0.792, respectively. A box and whisker plot showed higher 15-feature risk scores in the prevalent cases versus non-progressors (P < 0.0001, Fig. 2B). Logistic regression demonstrated that the 15-feature classifier could stratify patients with significantly different risks for prevalent HGD/EAC; ORs were 46.0 (95% CI, 14.86–169; P < 0.0001) for the comparison of the high-risk versus low-risk group and 7.67 (95% CI, 2.24–28.14; P = 0.001) for intermediate-risk versus low-risk group (Fig. 2C). The classifier identified both non-dysplastic and LGD biopsies from prevalent cases as high-risk (Fig. 2D). The probability of diagnosis of HGD/EAC on repeat endoscopy increased continuously as the 15-feature risk score increased (Fig. 2E). In multivariate logistic regression in which subsequent diagnosis of HGD/EAC was evaluated first in relation to pathologic diagnosis alone, then in relation to the predicted risk classes added to the pathologic diagnosis, the magnitude of ORs indicated that the predicted high-risk class provided independent and stronger predictive power than the generalist or gastrointestinal subspecialist pathologic diagnosis in this cohort of patients (Table 2). Similar results were obtained with multivariate Cox regression models (Supplementary Table S2).

The patients identified as high-risk exhibited multiple epithelial and stromal abnormalities that are quantified by the 15 image analysis features used by the risk classifier. Abnormalities detected in ND and LGD biopsies in patients with prevalent HGD/EAC included overexpression of p53, HER2/neu, and COX-2, focal AMACR overexpression, infiltration of the lamina propria by CD45RO-positive cells, CD68-positive cells, and stromal cells expressing HIF-1α (Fig. 3). The stronger predictive power of the risk classes compared with the pathologic diagnosis was illustrated in a patient with 2-cm segment Barrett's esophagus with biopsies available from two endoscopic levels. Biopsies from 32 and 34 cm were diagnosed as ND and LGD, respectively, by a gastrointestinal subspecialist in this study, and as ND and IND, respectively, by a general pathologist who recorded the original diagnosis. Repeat biopsy 56 days later showed HGD. Biopsies from the two levels, which were evaluated for illustrative purposes, scored 8.9 and 8.7 (on a scale of 0–10) with the 15-feature risk score, demonstrating that similar high-risk molecular and cellular changes were present at both biopsy levels despite the different pathologic diagnosis (Fig. 4A–J).

Using a case–control study design we validated a multivariable classifier that assesses ND, IND, and LGD biopsies to detect prevalent HGD/EAC in Barrett's esophagus patients. The test integrates quantitative biomarker and morphometric data into a risk score, and incorporates 3-tier risk stratification to classify patients as low-, intermediate-, or high-risk for HGD/EAC. The predicted high-risk group of patients was at 46-fold increased risk for prevalent HGD/EAC compared with the low-risk group. Importantly, the risk classes provided stronger predictive power than the expert GI and generalist pathologic diagnosis in this cohort of patients, and demonstrated high accuracy in detecting presence of prevalent HGD/EAC, even in nondysplastic biopsies. The tissue systems pathology assay used in this study thus has the potential to provide physicians and patients with an individualized score that indicates potential for prevalent HGD/EAC, which may aid in decision-making on more rigorous surveillance examinations and endoscopic therapy in Barrett's esophagus patients with ND, IND, or LGD. This study was strengthened by the use of a diverse patient cohort from four high-volume institutions in the United States and Europe. The study was further strengthened by the assay technology, which evaluates multiple pathways associated with carcinogenesis, and is also quantitative and objective. The assay can be performed on sections from FFPE blocks, and is thus compatible with clinical practice.

Expert referral centers have higher rates of detection of HGD/EAC and mucosal abnormalities than community centers, but recognition of subtle lesions containing HGD and early EAC can be challenging in all settings (7, 23). Endoscopic surveillance is effective when done in accordance with practice guidelines; however, adherence to the guidelines varies between settings (23, 24). HGD and early EAC may be missed by random sampling, which can result in repeat endoscopies and delays in diagnosis and treatment of HGD and EAC. A diagnosis of LGD confirmed by multiple gastrointestinal subspecialists is a strong predictor of malignant progression (25, 26). However, intraobserver variability in the diagnosis, even among gastrointestinal subspecialists, has been well documented (6, 27).

Despite extensive efforts to identify and validate biomarkers in Barrett's esophagus (13–16) none have yet been translated into practice to overcome the limitations of random sampling via detection of a field effect. p53 IHC has been demonstrated to have diagnostic and prognostic significance in Barrett's esophagus (17, 28). However, assessment of p53 alone is not sufficient since not all patients have detectable abnormalities in p53 protein levels, and a subset of patients who exhibit p53 abnormalities do not develop HGD or EAC (14, 29). Molecular approaches such as DNA sequencing, gene expression, mutation, and methylation profiling have been applied to diagnostic and prognostic testing in Barrett's esophagus (30–32) but have not yet been evaluated in the detection of abnormalities in the expanded field surrounding HGD and EAC. These technologies have the disadvantage of requiring tissue digestion, resulting in loss of contextual information, such as nuclear morphology, and spatial relationships that are relevant to patient outcomes. Further disadvantages include the requirement for fresh frozen specimens for some of these genomic approaches, which is a logistical problem in clinical practice, and also the need for laser microdissection of tissue areas based on subjective review for some of these approaches.

The risk classifier evaluated in this study identifies patients who have prevalent HGD/EAC, despite receiving a pathologic diagnosis of ND, IND, or LGD. The abnormalities quantified by the assay include loss of tumor suppression, loss of cell-cycle control, morphologic changes, increased inflammation, stromal angiogenesis, and altered patterns of infiltrating immune cells (20). If validated in additional studies, this finding suggests that objective detection of multiple molecular and cellular abnormalities in the preneoplastic field of Barrett's esophagus could overcome some of the limitations of random endoscopic sampling and pathologic diagnosis, and enable earlier detection of HGD/EAC in all practice settings. Additional studies are required to evaluate the performance of the assay in biopsies rom different endoscopic levels. Ex vivo tests such as this may also complement newer endoscopic techniques such as volumetric laser endomicroscopy (33) by providing objective, quantitative analysis of pathways involved in malignant transformation.

We readily recognize the limitations of this study, which include the retrospective nature of the study, the case–control cohort study design in which the proportion of prevalent cases was not representative of the general population, and the small number of available prevalent cases as reflected in the wide confidence intervals we report. Additionally, larger studies will be required to validate our findings. However, large prospective studies are challenging in Barrett's esophagus due to the low prevalence of malignant progression. The set of biopsies and patients was heterogeneous; the biopsies had diagnoses of ND, IND, and LGD, 22 of 30 patients had prevalent HGD whereas 8 of 30 had prevalent EAC, and the intervals between the biopsy tested and the repeat endoscopy demonstrating HGD/EAC were variable. Biopsies from a single endoscopy level were evaluated in this study, and additional studies are required with biopsies from multiple endoscopy levels from patients with subtle lesions containing HGD or early EAC. The cohort in this study included patients in surveillance at multiple centers in the United States and Europe over a wide timeframe, which prevented standardization of preanalytic variables. However, the biopsies reflect routine samples requiring accurate risk assessment.

The assay requires instrumentation and software that could not easily be integrated into current pathology laboratories, which are only beginning to adopt digital pathology. The assay has been deployed in a central reference laboratory equipped with the necessary resources, which enables physicians at expert referral centers and community centers to order the assay. Unstained slides can be sent to the reference laboratory, the assay is performed and the laboratory provides a clinical report to the ordering physician and submitting pathologist. The testing process (Supplementary Fig. S1) takes approximately 3 business days. The testing approach would initially add to the cost of Barrett's esophagus surveillance. However, risk prediction testing has the potential to result in cost savings in high-risk patients by reducing repeat endoscopies and pathologist time required to diagnose HGD/EAC, enabling earlier intervention with endoscopic therapies to reduce EAC incidence and mortality. In low-risk patients there is the potential to lower future costs by extending surveillance intervals (34, 35). Although it is feasible to test biopsies from multiple levels in long segment Barrett's esophagus, this may add significant cost in a subset of patients, which may outweigh the potential cost savings from reducing unnecessary endoscopies or intervening early to prevent progression.

In summary, the tissue systems pathology assay examined in this study objectively quantifies multiple epithelial and stromal processes that predict prevalent HGD/EAC in Barrett's esophagus patients. The assay has the potential to improve upon current histology methods to enable earlier detection of HGD/EAC, which will facilitate earlier, more effective therapeutic interventions.

R.J. Critchley-Thorne has ownership interest (including patents) in Cernostics, Inc. and has provided expert testimony as an inventor on patents and patent applications for the TissueCypher technology used in this study. J.W. Prichard, J. Li, and D.L. Diehl provided expert testimony for primary employment at Geisinger Health System, which has a financial interest in Cernostics, Inc., the commercial entity that developed the proprietary TissueCypher technology used in this study. L.M. Reese is a senior associate scientist and has ownership interest (including patents) in Cernostics. K.A. Repa has provided expert testimony as a former full-time employee of Cernostics, Inc. J.J.G.H.M. Bergman reports receiving commercial research support from Cernostics. Y. Zhang is a consultant/advisory board member for Cernostics. A. Rustgi has ownership interest in Cernostics, Inc. No potential conflicts of interest were disclosed by the other authors.

Conception and design: R.J. Critchley-Thorne, J.W. Prichard, J. Li, B.A. Jobe, A.H. Zaidi, J.J.G.H.M. Bergman, A. Rustgi

Development of methodology: R.J. Critchley-Thorne, J.M. Davison, K. Repa, J. Li, J.J.G.H.M. Bergman, A. Rustgi

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): R.J. Critchley-Thorne, J.M. Davison, L.M. Reese, K. Repa, J. Li, G.G. Ginsberg, M. DeMarshall, T. Foxwell, L.C. Duits, J.J.G.H.M. Bergman, G.W. Falk

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): R.J. Critchley-Thorne, J.M. Davison, L.M. Reese, Y. Zhang, J. Li, D.L. Diehl, N.C. Jhala, B.A. Jobe, A.H. Zaidi, J.J.G.H.M. Bergman, A. Rustgi

Writing, review, and/or revision of the manuscript: R.J. Critchley-Thorne, J.M. Davison, J.W. Prichard, L.M. Reese, Y. Zhang, J. Li, D.L. Diehl, N.C. Jhala, B.A. Jobe, A.H. Zaidi, L.C. Duits, J.J.G.H.M. Bergman, A. Rustgi, G.W. Falk

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): R.J. Critchley-Thorne, L.M. Reese, K. Repa, L.C. Duits

Study supervision: R.J. Critchley-Thorne, J.J.G.H.M. Bergman

We thank Georgiann Noll and Matthew Barley (Geisinger) for assistance with specimen retrieval, and D. Lansing Taylor, PhD; Janet Warrington, PhD; and Aaron DeWard, PhD, for many helpful discussions and advice on the research study.

This work was supported by a grant from the Pennsylvania Department of Health Cure Program Grant, Research on Cancer Diagnostics or Therapeutics with Commercialization Potential (RFA#10-07-03; to R.J. Critchley-Thorne, J.M. Davison, G.W. Falk, J.W. Prichard, and Y. Zhang), a Qualifying Therapeutic Discovery Project Grant, Internal Revenue Service/Affordable Care Act 2010 (to R.J. Critchley-Thorne), and by Cernostics, Inc. Partial support was provided by the National Cancer Institute of the NIH under Award number R44CA192416 (to R.J. Critchley-Thorne and J.M. Davison).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.