TUM Tests 19 Pathology Foundation Models: Cancer Found, Dysplasia Still Hard
Researchers Kazemi, Schüffler et al. at the Technical University of Munich (TUM) published a paper in npj Digital Medicine titled "Nineteen Foundation Models and Barrett's Esophagus: Cancer Is Found, Dysplasia Is Still Hard." Using the BarrettNet multi-center cohort (21 centers, 790 patients, 3,528 whole-slide images), they systematically benchmarked contemporary pathology foundation models on Barrett's esophagus diagnosis while simulating real-world class imbalance. Their conclusion: cancer can be detected, but dysplasia remains a tough nut to crack.
Confirmed
- The study covers the full spectrum of Barrett's esophagus (from normal tissue to cancer) and involves 19 pathology foundation models
- Pathology foundation models clearly outperform the ResNet50 baseline, with AUROC around 0.90 vs. 0.81
- Only 52% of the 145 dysplasia slides were correctly identified as dysplasia—and dysplasia is precisely the stage where pathologists disagree most
- An ensemble strategy averaging the top five models yielded little gain: grading AUROC improved only from 0.901 to 0.907
Unconfirmed
- Limitations noted by the authors themselves: PAXgene tissue processing, a single cohort, and lack of external validation
- The ensemble models were selected within the same cross-validation used for evaluation, so the real-world generalization of their slight improvement is questionable
Why it matters
The study shows that pathology foundation models are already quite capable of detecting outright cancer, but remain unreliable at the precancerous (dysplasia) stage—the hardest and most clinically urgent area where decision support is needed. Deploying AI in real pathology workflows will require improving models for contested diagnostic scenarios and validating them through independent external studies.
2026-10-27 ~ 2026-10-27 · 5 related posts
Primary sources
- [source] New paper: 19 foundation models tested on Barrett's esophagus diagnosis — maier_ak · 2026-10-27
- [source] TUM study: 19 pathology foundation models find cancer but miss dysplasia 48% of the time — maier_ak · 2026-10-27
- [source] Only 52% of dysplastic slides recognized, TUM pathology study finds — maier_ak · 2026-10-27
- Pathology foundation models hit AUROC 0.90 vs ResNet50's 0.81 — maier_ak · 2026-10-27
1 near-duplicate retellings: maier_ak