Open-source CT tumor AI: 5 false positives in 80 healthy scans, catches 33 of 52 adrenal cancers

MaziyarPanahi · x · 2026-11-03

The author stress-tested his CT tumor-detection tool on 80 healthy abdominal CTs using one GPU: at a 1 mL threshold it flagged 5 false positives while still detecting 33 of 52 adrenal cancers, prompting radiologists to weigh in on how small a lesion AI should report.

He also open-sourced Merlin-Cancer-Super (Johns Hopkins + Stanford collaboration):

Based on R-Super (MICCAI 2025 Best Paper), which uses an LLM to extract tumor info from radiology reports and converts it into per-voxel supervision via new Volume Loss and Ball Loss functions
Surpasses public AI models and leading VLMs at detecting/segmenting tumors across 9 organs (bladder, gallbladder, spleen, esophagus, stomach, duodenum, prostate, uterus, adrenals)
Trained on Stanford's Merlin Plus dataset (25,494 CT scans); author cautions possible overlap with NIH scans and that healthy/cancer scans come from different hospitals

Weights and dataset are available on Hugging Face under CC BY-NC-ND.

Original post →

Open-source CT tumor AI: 5 false positives in 80 healthy scans, catches 33 of 52 adrenal cancers

More from Models