A deep learning system for detection of early Barrett's neoplasia: a model development and validation study.

BACKGROUND: Computer-aided detection (CADe) systems could assist endoscopists in detecting early neoplasia in Barrett's oesophagus, which could be difficult to detect in endoscopic images. The aim of this study was to develop, test, and benchmark a CADe system for early neoplasia in Barrett's oesophagus. METHODS: The CADe system was first pretrained with ImageNet followed by domain-specific pretraining with GastroNet. We trained the CADe system on a dataset of 14 046 images (2506 patients) of confirmed Barrett's oesophagus neoplasia and non-dysplastic Barrett's oesophagus from 15 centres. Neoplasia was delineated by 14 Barrett's oesophagus experts for all datasets. We tested the performance of the CADe system on two independent test sets. The all-comers test set comprised 327 (73 patients) non-dysplastic Barrett's oesophagus images, 82 (46 patients) neoplastic images, 180 (66 of the same patients) non-dysplastic Barrett's oesophagus videos, and 71 (45 of the same patients) neoplastic videos. The benchmarking test set comprised 100 (50 patients) neoplastic images, 300 (125 patients) non-dysplastic images, 47 (47 of the same patients) neoplastic videos, and 141 (82 of the same patients) non-dysplastic videos, and was enriched with subtle neoplasia cases. The benchmarking test set was evaluated by 112 endoscopists from six countries (first without CADe and, after 6 weeks, with CADe) and by 28 external international Barrett's oesophagus experts. The primary outcome was the sensitivity of Barrett's neoplasia detection by general endoscopists without CADe assistance versus with CADe assistance on the benchmarking test set. We compared sensitivity using a mixed-effects logistic regression model with conditional odds ratios (ORs; likelihood profile 95% CIs). FINDINGS: Sensitivity for neoplasia detection among endoscopists increased from 74% to 88% with CADe assistance (OR 2·04; 95% CI 1·73-2·42; p<0·0001 for images and from 67% to 79% [2·35; 1·90-2·94; p<0·0001] for video) without compromising specificity (from 89% to 90% [1·07; 0·96-1·19; p=0·20] for images and from 96% to 94% [0·94; 0·79-1·11; ] for video; p=0·46). In the all-comers test set, CADe detected neoplastic lesions in 95% (88-98) of images and 97% (90-99) of videos. In the benchmarking test set, the CADe system was superior to endoscopists in detecting neoplasia (90% vs 74% [OR 3·75; 95% CI 1·93-8·05; p=0·0002] for images and 91% vs 67% [11·68; 3·85-47·53; p<0·0001] for video) and non-inferior to Barrett's oesophagus experts (90% vs 87% [OR 1·74; 95% CI 0·83-3·65] for images and 91% vs 86% [2·94; 0·99-11·40] for video). INTERPRETATION: CADe outperformed endoscopists in detecting Barrett's oesophagus neoplasia and, when used as an assistive tool, it improved their detection rate. CADe detected virtually all neoplasia in a test set of consecutive cases. FUNDING: Olympus.

The SpaTemp cohort: 168 nondysplastic Barrett's esophagus surveillance patients with and without progression to early neoplasia to evaluate the distribution of biomarkers over space and time.

The ReBus cohort is a matched nested case-control cohort of patients with nondysplastic (ND) Barrett's esophagus (BE) at baseline who progressed (progressors) or did not progress (nonprogressors) to high-grade dysplasia (HGD) or cancer. This cohort is constructed using the most stringent inclusion criteria to optimize explorative studies on biomarkers predicting malignant progression in NDBE. These explorative studies may benefit from expanding the number of cases and by incorporating samples that allow assessment of the biomarker over space (spatial variability) and over time (temporal variability). To (i) update the ReBus cohort by identifying new progressors and (ii) identify progressors and nonprogressors within the updated ReBus cohort containing spatial and temporal information. The ReBus cohort was updated by identifying Barrett's patients referred for endoscopic work-up of neoplasia at 4 tertiary referral centers. Progressors and nonprogressors with a multilevel (spatial) endoscopy and additional prior (temporal) endoscopies were identified to evaluate biomarkers over space and over time. The original ReBus cohort consisted of 165 progressors and 723 nonprogressors. We identified 65 new progressors meeting the same strict selection criteria, resulting in a total number of 230 progressors and 723 matched nonprogressors in the updated ReBus cohort. Within the updated cohort, 61 progressors and 107 nonprogressors (mean age 61 ± 10 years) with a spatial endoscopy (median level 3 [2-4]) were identified. 33/61 progressors and 50/107 nonprogressors had a median of 3 (2-4) additional temporal endoscopies. Our updated ReBus cohort consists of 230 progressors and 723 matched nonprogressors using the most strict selection criteria. In a subgroup of 168 Barrett's patients (the SpaTemp cohort), multiple levels have been sampled at baseline and during follow-up providing a unique platform to study spatial and temporal distribution of biomarkers in BE.

Adherence to pre-set benchmark quality criteria to qualify as expert assessor of dysplasia in Barrett's esophagus biopsies - towards digital review of Barrett's esophagus.

Background: Dysplasia assessment of Barrett's esophagus biopsies is associated with low observer agreement; guidelines advise expert review. We have developed a web-based review panel for dysplastic Barrett's esophagus biopsies. Objective: The purpose of this study was to test if 10 gastrointestinal pathologists working at Dutch Barrett's esophagus expert centres met pre-set benchmark scores for quality criteria. Methods: Ten gastrointestinal pathologists twice assessed 60 digitalized Barrett's esophagus cases, enriched for dysplasia; then randomised (7520 assessments). We tested predefined benchmark quality criteria: (a) percentage of 'indefinite for dysplasia' diagnoses, benchmark score ≤14% for all cases, ≤16% for dysplastic subset, (b) intra-observer agreement; benchmark score ≥0.66/≥0.39, (c) percentage agreement with 'gold standard diagnosis'; benchmark score ≥82%/≥73%, (d) proportion of cases with high-grade dysplasia underdiagnosed as non-dysplastic Barrett's esophagus; benchmark score ≤1/78 (≤1.28%) assessments for dysplastic subset. Results: Gastrointestinal pathologists had seven years' Barrett's esophagus-experience, handling seven Barrett's esophagus-cases weekly. Three met stringent benchmark scores; all cases and dysplastic subset, three met extended benchmark scores. Four pathologists lacked one quality criterion to meet benchmark scores. Conclusion: Predefined benchmark scores for expert assessment of Barrett's esophagus dysplasia biopsies are stringent and met by some gastrointestinal pathologists. The majority of assessors however, only showed limited deviation from benchmark scores. We expect further training with group discussions will lead to adherence of all participating gastrointestinal pathologists to quality criteria, and therefore eligible to join the review panel.

The Amsterdam ReBus progressor cohort: identification of 165 Barrett's surveillance patients who progressed to early neoplasia and 723 nonprogressor patients.

Patient selection is suboptimal in most studies focused on identifying biological markers for neoplastic progression in Barrett's esophagus (BE). This study aims to describe a stringently selected community-based case-control cohort of non-dysplastic BE (NDBE) patients who progressed to high-grade dysplasia (HGD) or esophageal adenocarcinoma (EAC) and BE patients who never progressed to be used for future biomarker studies. We identified all patients referred for endoscopic work-up of BE neoplasia at three tertiary referral centers for treatment of BE neoplasia between 2000 and 2013. We performed a detailed registration of any endoscopic surveillance history before neoplastic progression. Controls were selected from a retrospective BE surveillance registration in 10 community hospitals. A total of 887 patients were referred for endoscopic work-up of BE neoplasia. Based on predefined selection criteria, we identified 165 progressor patients (82% men; mean age 55 years ± 10.4) with a baseline endoscopy demonstrating NDBE > 2 years before neoplastic progression. Using the same predefined selection criteria, 723 nonprogressor patients (67% men; mean age 57 years ± 11.3) with >2 years of endoscopic surveillance were identified. Median length of the BE segment was 5 cm (IQR 4-7) in progressors and 4 cm (IQR 2-6) in controls. Median duration of surveillance was 89 months (IQR 54-139) in progressors and 76 months (IQR 47-116) in nonprogressors. Paraffin embedded biopsies are available for biomarker research in all patients. Ethical approval was obtained and material transfer agreements were signed with all 58 contributing pathology labs. This is the largest community-based case-control cohort of BE patients with and without progression to early neoplasia. The stringent selection criteria and the availability of paraffin embedded biopsy specimens make this a unique cohort for biomarker studies.

A biomarker panel predicts progression of Barrett's esophagus to esophageal adenocarcinoma.

Progression from Barrett's esophagus (BE) to esophageal adenocarcinoma (EAC) is uncommon but the consequences are serious. Predictors of progression are essential to optimize resource utilization. This study assessed the utility of a promising panel of biomarkers applicable to routine paraffin embedded biopsies (FFPE) to predict progression of BE to EAC in a large population-based, nested case-control study.We utilized the Amsterdam-based ReBus nested case-control cohort. BE patients who progressed to high-grade dysplasia (HGD)/EAC (n = 130) and BE patients who never progressed (n = 130) were matched on age, sex, length of the BE segment, and duration of endoscopic surveillance. All progressors had minimum 2 years of endoscopic surveillance without HGD/EAC to exclude prevalent neoplasia. We assessed abnormal DNA content, p53, Cyclin A, and Aspergillus oryzae lectin (AOL) in FFPE sections. We performed conditional logistic regression analysis to estimate odds ratio (OR) of progression based on biomarker status.Expert LGD (OR, 8.3; 95% CI, 1.7-41.0), AOL (3 vs. 0 epithelial compartments abnormal; OR, 3.6; 95% CI, 1.2-10.6) and p53 (OR, 2.3; 95% CI, 1.2-4.6) were independently associated with neoplastic progression. Cyclin A did not predict progression and DNA ploidy analysis by image cytometry was unsuccessful in the majority of cases, both were excluded from the multivariate analysis. The multivariable biomarker model had an area under the receiver operating characteristic curve of 0.73.Expert LGD, AOL, and p53 independently predict neoplastic progression in BE patients and are applicable to routine practice. These biomarkers can aid in selecting patients for endoscopic ablation or more intensive surveillance.

Development of benchmark quality criteria for assessing whole-endoscopy Barrett's esophagus biopsy cases.

BACKGROUND: Dysplasia in Barrett's esophagus (BE) biopsies is associated with low observer agreement among general pathologists. Therefore, expert review is advised. We are developing a web-based, national expert review panel for histological review of BE biopsies. OBJECTIVE: The aim of this study was to create benchmark quality criteria for future members. METHODS: Five expert BE pathologists, with 10-30 years of BE experience, weekly handling 5-10 cases (25% dysplastic), assessed a case set of 60 digitalized cases, enriched for dysplasia. Each case contained all slides from one endoscopy (non-dysplastic BE (NDBE), n = 21; low-grade dysplasia (LGD), n = 20; high-grade dysplasia (HGD), n = 19). All cases were randomized and assessed twice followed by group discussions to create a consensus diagnosis. Outcome measures: percentage of 'indefinite for dysplasia' (IND) diagnoses, intra-observer agreement, and agreement with the consensus 'gold standard' diagnosis. RESULTS: Mean percentage of IND diagnoses was 8% (3-14%) and mean intra-observer agreement was 0.84 (0.66-1.02). Mean agreement with the consensus diagnosis was 90% (95% prediction interval (PI) 82-98%). CONCLUSION: Expert pathology review of BE requires the scoring of a limited number of IND cases, consistency of assessment and a high agreement with a consensus gold standard diagnosis. These benchmark quality criteria will be used to assess the performance of other pathologists joining our panel.

Validation of a biomarker panel in Barrett's esophagus to predict progression to esophageal adenocarcinoma.

In a prior study, baseline mutational load (ML) predicted progression to high-grade dysplasia (HGD) or esophageal adenocarcinoma (EAC) in Barrett's esophagus (BE) with an area under the curve (AUC) of 0.95. We aimed to validate the test characteristics of this predictive biomarker panel using crude DNA lysates in a larger well-characterized cohort. We performed a nested case-control study of BE patients from three tertiary referral centers in the Netherlands. Cases had baseline nondysplastic BE (NDBE) and developed HGD/EAC ≥ 2 years later. Controls were matched 2:1, had baseline NDBE, and no progression. Polymerase chain reaction (PCR)-based mutational analysis was performed on crude lysates from formalin-fixed, paraffin-embedded tissue. ML was calculated from loss of heterozygosity (LOH) and microsatellite instability (MSI) at 10 genomic loci. Receiver operator characteristic (ROC) curves were created to assess the diagnostic utility of various cutoffs of ML for progression. Of 159 subjects, 58 were progressors and 101 were nonprogressors, there was no difference in mean ML in preprogression tissue in progressors and nonprogressors (ML = 0.73 ± 0.69 vs. ML = 0.74 ± 0.61, P = 0.93). ROC curves showed poor discrimination of ML in predicting progression with AUC of 0.50 at ML ≥ 1. AUC did not vary with different ML cut-points. The utility of the ML to stratify BE patients for risk of progression was not confirmed in this study. The etiology for discrepancies between this and prior studies showing high predictiveness is likely due to the use of crude lysates in this study, but requires further investigation.

Digital microscopy as valid alternative to conventional microscopy for histological evaluation of Barrett's esophagus biopsies.

Management of Barrett's esophagus (BE) relies heavily on histopathological assessment of biopsies, associated with significant intra- and interobserver variability. Guidelines recommend biopsy review by an expert in case of dysplasia. Conventional review of biopsies, however, is impractical and does not allow for teleconferencing or annotations. An expert digital review platform might overcome these limitations. We compared diagnostic agreement of digital and conventional microscopy for diagnosing BE ± dysplasia. Sixty BE biopsy glass slides (non-dysplastic BE (NDBE); n = 25, low-grade dysplasia (LGD); n = 20; high-grade dysplasia (HGD); n = 15) were scanned at ×20 magnification. The slides were assessed four times by five expert BE pathologists, all practicing histopathologists (range: 5-30 years), in 2 alternating rounds of digital and conventional microscopy, each in randomized order and sequence of slides. Intraobserver and pairwise interobserver agreement were calculated, using custom weighted Cohen's kappa, adjusted for the maximum possible kappa scores. Split into three categories (NDBE, IND, LGD+HGD), the mean intraobserver agreement was 0.75 and 0.84 for digital and conventional assessment, respectively (p = 0.35). Mean pairwise interobserver agreement was 0.80 for digital and 0.85 for conventional microscopy (p = 0.17). In 47/60 (78%) of digital microscopy reviews a majority vote of ≥3 pathologists was reached before consensus meeting. After group discussion, a majority vote was achieved in all cases (60/60). Diagnostic agreement of digital microscopy is comparable to that of conventional microscopy. These outcomes justify the use of digital slides in a nationwide, web-based BE revision platform in the Netherlands. This will overcome the practical issues associated with conventional histologic review by multiple pathologists.