Alibaba open-sources Damo Radar, a CT-reading AI model that beat 23 of 26 radiologists in a Science study
- Alibaba's research arm Damo Academy open-sourced Damo Radar, a vision-language model that reads contrast-enhanced CT scans of 18 abdominal organs and flags 146 clinical findings, including cancers.
- Across nearly 40,000 real-world exams the model averaged an AUC of 0.913 over those 146 findings, where 1.0 is perfect diagnostic accuracy.
- In a comparison with 26 radiologists from multiple hospitals, the model's average accuracy beat 23 of them; radiologists working with it missed 10% fewer diagnoses and cut reading time by more than 30%.
- The study ran in Science on Thursday with a Zhejiang University-affiliated hospital among the participating institutions, and the model was trained on CT scans paired with clinical reports.
- It follows Alibaba's April release of the Coca AI model for early colorectal cancer detection, and the team calls Damo Radar the first expert-level generalist medical imaging model.
Hacker News opinions
Anyone know how many parameters this thing has? Nobody in the coverage says.
Weights are only about 5GB, so this is a small model. Kind of remarkable for those numbers.
The actual paper is up: 'Large-scale AI-guided liver malignancy diagnosis: multicenter study and a single-arm trial' in Nature Medicine. Weights on Hugging Face at radar-generalist, code in the damo-academy repo. Note it's liver-focused, unlike the SCMP headline.
This fits a pattern I keep collecting: a Gemma model helping find a new cancer therapy pathway, exopolysaccharide EPS3.9 triggering pyroptosis in mice with liver cancer, eCPMV virus nanoparticles. The renal stuff is a different thread.
Honestly I mostly ask Gemini medical questions and it tells me I'm very smart and almost always right, which is exactly what I want from a doctor.