Open-source Merlin model from JHU+Stanford auto-segments tumors on CT in one click
MaziyarPanahi · x · 2026-10-27
Maziyar Panahi fed a CT scan from a patient whose adrenal cancer was found by chance to Merlin-Cancer-Super, an open model from Johns Hopkins + Stanford. No hints, one click, one GPU — and the model outlined the tumor right where the radiologist did.
Key details:
- Built on R-Super (Report Supervision), MICCAI 2025 Best Paper Award (top 2): an LLM extracts tumor info from radiology reports, converted into per-voxel supervision via novel Volume Loss & Ball Loss constraints on tumor count, diameter, and location
- Trained on Merlin Plus: longitudinal metadata and per-voxel annotations for organs and 9 tumor types across 25,494 Stanford CT scans
- Outperforms prior public AI models and leading VLMs at tumor detection/segmentation across 9 organs (bladder, gallbladder, spleen, esophagus, stomach, duodenum, prostate, uterus, adrenals)
Released under CC BY-NC-ND 4.0 on Hugging Face.
Related event: Open Model from JHU and Stanford Detects Tumor in CT Scan(2 posts)→
More from Apps
- ChatGPT Chrome extension hides the full website: drag it left to reveal sidebar — mark_k · 2026-10-27
- 10 open-source GitHub repos to build a local-first AI second brain — victor_explore · 2026-10-27
- NewMax runs optimized Claude Code with BYOK keys, plus pool rotation and OAuth — yangyi · 2026-10-27
- He built a 90s cable TV app for Plex with Claude: 1,400+ downloads and open source — chadMueller · 2026-10-27
- How to Get Cited by ChatGPT: A 6-Step AEO Playbook — jamesbrooksco · 2026-10-27
- How a site went from zero to most traffic from ChatGPT: an AI-search optimization checklist — jamesbrooksco · 2026-10-27