Open-source cancer model hits tumors in 34/34 CTs but median Dice is only 0.22

MaziyarPanahi · x · 2026-11-20

Maziyar Panahi ran the open-source Merlin-Cancer-Super model (Johns Hopkins + Stanford) on 34 public CTs of women with uterine cancer, no hints: it touched every tumor (34/34), a single scan took 107s on one A100, and the radiologists' outlines never went to the model. Best case hit Dice 0.83 inside the radiologist's outline, but it drew 30 of 34 tumors smaller than the doctors did, median Dice is just 0.22, and the authors' 8.7 mL cutoff flagged 17 cases. Impressive recall, far from clinical segmentation accuracy; no false-alarm rate measurable yet since every scan had cancer.

Related event: Open-Source Cancer Model Detects Tumors in All 34 Uterine Cancer CTs(3 posts)→

Original post →

More from Models

Models channel →