Paper review · November 21, 2026
Multi-center grading of Barrett's esophagus dysplasia without stain normalization, using a pathology foundation model and MIL
Multi-institutional grading of Barrett's esophagus dysplasia without stain normalization using a pathology foundation model and multiple instance learning
Using pathology foundation model features and two-stage MIL to classify Barrett's esophagus dysplasia into three grades, the study achieved a macro AUC of 0.96 on data from four institutions without stain normalization.
Problem
The dysplasia grade in Barrett's esophagus is the basis for deciding treatment. However, interobserver agreement among pathologists is low, especially for low-grade dysplasia (κ=0.42). Existing deep learning systems needed a stain normalization model for each institution, which made broad deployment across many hospitals difficult.
Method
The study used 969 H&E whole slide images from four institutions. 1280×1280 patches were cut from the slides, and features were extracted with Midnight-12k, a pathology-specific foundation model. In stage 1, four binary MLP classifiers (NDBE, dysplasia, LGD, HGD), trained on expert annotations from Mayo Clinic, classify the patches. Based on these results, priority sampling builds a bag of 1,000 patches per slide. In stage 2, AttriMIL combines class-specific attention with patch scores to classify each slide as NDBE, LGD or HGD. No stain normalization was applied.
Results
- On a test set of 130 slides, the model recorded a macro AUC of 0.96 (95% CI 0.93–0.98) and a macro F1 of 0.85. Per-class AUC was 0.94 for LGD and 0.97 each for NDBE and HGD.
- For the binary call of dysplasia present or absent, sensitivity was 98% (80/82), specificity was 81% (39/48), and the negative likelihood ratio was 0.030.
- Quadratic-weighted κ against expert consensus was 0.83. No HGD case was misclassified as NDBE, and 14 of the 20 misclassifications assigned a higher grade than the true one.
- Macro AUC by institution ranged from 0.94 to 0.96, and differences in accuracy between institutions were not significant (Fisher p=0.34).
Limitations
This is a retrospective study, and the test set is small at 130 slides. Per-institution samples are very small, such as 21 slides from Baylor (3 HGD), so comparisons between institutions have low statistical power. Patch-level annotations existed only for Mayo Clinic slides. With 81% specificity, the model tends to overgrade NDBE. No experiment directly comparing against a setup with stain normalization appears in the excerpt. Evaluation used a test set split from the same four institutions rather than validation at an independent external institution.
Why it matters
The study provides evidence that using a foundation model's stain-robust features can enable consistent pathology reading support across multiple hospitals without institution-specific preprocessing. Given its high negative predictive power, it is worth examining whether it could serve as a screening tool for ruling out dysplasia.
Key terms
- MIL(Multiple Instance Learning)
- A method that, when only slide-level labels exist and individual patches are unlabeled, takes a bag of patches as input and learns slide-level predictions.
- Pathology foundation model
- A vision model pretrained with self-supervision on a large number of pathology slides. This study used Midnight-12k, trained with the DINOv2 approach.
- Stain normalization
- A preprocessing step that standardizes H&E stain coloring, which varies by institution. This study skipped this step.
- Negative likelihood ratio (Negative LR)
- A measure of how much a negative test result lowers the likelihood of disease. The smaller the value, the more trustworthy a negative result.
Original
Authors: Amirali Khosravi, Jaidip M. Jagtap, D. Chamil Codipilly, Catherine E. Hagen, Jason T. Lewis, Rickey E. Carter, Shahriar Faghani, Vani J. A. Konda, et al. · Posted 2026-11-19 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.