TUM study: 19 pathology foundation models find cancer but miss dysplasia 48% of the time
maier_ak · x · 2026-10-27
A TUM study in npj Digital Medicine tested 19 pathology foundation models on 3,528 biopsy slides from 790 patients across the full Barrett's esophagus spectrum.
Key findings
- Foundation models clearly beat a ResNet50 baseline: AUROC 0.90 vs 0.81.
- Only 52% of dysplastic slides (just 145) are recognized as dysplasia — exactly the stage where pathologists disagree most.
- Averaging the best five models adds little: 0.901 → 0.907 for subtyping, selected on the same cross-validation used for evaluation.
- Limitations: PAXgene tissue, single cohort, no external validation.
Bottom line: great at finding cancer, unreliable at the hardest pre-cancerous call.
Related event: TUM Benchmarks 19 Pathology Foundation Models on Barrett's Esophagus(5 posts)→
More from Research
- S3AC: Shanghai AI Lab's Stabilized Actor-Critic Method for LLM Long-Chain Reasoning — Shanghai-AI-Laboratory · 2026-10-27
- Skill∞: Recursive Self-Improvement Where Agents Optimize Their Own Optimization Skills — NationalUniversityofSingapore · 2026-10-27
- 229k-sample replication of Cohere's Qwen 3.8 malicious PUT claim raises more questions — builderjaydub · 2026-10-27
- AFROSCOPE identifies 640 African languages, beats frontier LLMs' 47.4 F1 with 97.9 — mageed · 2026-10-27
- Physicist on OpenAI's math result: it's mathematical physics, not theoretical physics — skdh · 2026-10-27
- Mechanistic Interpretability and Fast Memory Recall Are the Same Problem — mayfer · 2026-10-27