19 pathology foundation models tested: cancer is found, dysplasia is still hard
maier_ak · x · 2026-10-27
A study by Kazemi, Schüffler et al. (TU Munich) in npj Digital Medicine benchmarks current pathology foundation models on Barrett's esophagus surveillance using BarrettNet, a German multicentre cohort: 21 centres, 790 patients, 3,528 whole-slide images.
- The set spans the full real-world spectrum — 41% normal, 48% non-dysplastic Barrett's, 4% dysplastic, 7% cancer — not a tidy two-class benchmark.
- Key number: only 52% of dysplastic slides are recognized as dysplasia (on just 145 slides), exactly the grading stage where expert pathologists disagree most; cancer detection performs far better.
- Caveats: PAXgene tissue processing, one cohort, no external validation.
Takeaway: AI finds cancer, but dysplasia grading remains the weak spot of pathology AI.
More from Research
- Dev quips: if 3SAT constants are cosmic, P=NP wouldn't matter anyway — sytelus · 2026-10-27
- Bittensor subnet SN73 rebrands to Elenost, benchmarking perception in a simulated city — markjeffrey · 2026-10-27
- A 30-minute workflow for triaging any research paper — cneuralnetwork · 2026-10-27
- Nature Biotechnology: AI should join peer review as a bounded, non-voting, auditable 'reviewer' — OdedRechavi · 2026-10-27
- Ablation table for Game24-from-Countdown shows hierarchical RL hitting 69.7 vs flat baseline 52.4 — vxnuaj · 2026-10-27
- Google maps every hedgerow in England with deep learning, making hidden ecology visible — danielrock · 2026-10-27