Alibaba open-sources RADAR: one CT scan, 146 findings, Apache-2.0
DAMO Academy's model out-read 23 of 26 radiologists at 0.913 AUC, cut reading time 30%, and the weights are downloadable today — a rare open medical unlock.
Copy markdown
146 findings, 18 organs, one scan
RADAR, from Alibaba's DAMO Academy and Zhejiang University, reads a single contrast-enhanced abdominal CT and flags 146 findings — cancers, lesions, and other abnormalities — across 18 organs. It trained on 424,911 CT exams and 15M+ anatomy-aware image-text pairs, and the work ran in Science.
It out-read 23 of 26 radiologists
Mean AUC hit 0.913 across ~40,000 real-world exams, held at 0.895 across eight outside hospitals, and 0.904 on 27,000+ emergency scans. As a second reader it lifted radiologists' sensitivity ~10% and cut reading time by over 30%.
Apache-2.0, weights up now
Code, weights, and download scripts are live on GitHub and Hugging Face under Apache-2.0 (a few components carry non-commercial terms, so check before shipping). That makes RADAR one of the few Science-grade medical models you can actually pull down, fine-tune, and wrap in your own service.
What you could build this weekend
A triage layer that pre-flags urgent abdominal findings on a radiology worklist, or a research pipeline over open CT datasets with no per-inference API bill. The honest caveat: it's a research release, not FDA- or CE-cleared — keep a clinician in the loop and treat outputs as decision support, not diagnosis.