Alibaba's Damo Academy open-sources Damo Radar, a medical AI model detecting ~150 abdominal conditions from CT scans
Alibaba's research arm Damo Academy publicly released (open-sourced) a vision-language AI model, Damo Radar, trained to analyze contrast-enhanced CT scans of 18 abdominal organs and identify 146 clinical findings including cancers. A study on the model, evaluated on nearly 40,000 real-world exams with an average AUC of 0.913 across 146 findings, was published in Science.
Entities: Alibaba Group Holding, Damo Academy, Damo Radar, Science
0 primary
What happened
Alibaba's Damo Academy has open-sourced Damo Radar, a vision-language AI model trained to read contrast-enhanced CT scans across 18 abdominal organs and flag 146 possible findings, including cancers. A companion study published in Science reports an average AUC of 0.913 across those 146 findings, tested against nearly 40,000 real-world exams. The reporting here is secondary (via Hacker News/SCMP); no links to the paper, code repository or model weights are provided in the source material reviewed.
Why it matters
If the release holds up, it puts a radiology-grade diagnostic model into the hands of any hospital, researcher or startup with the compute to run it, rather than behind a vendor licence. That matters most for lower-resource health systems and AI developers who can now build on a benchmarked base model instead of starting from scratch. The catch is that open-sourcing a model is not the same as deploying it: nothing here addresses regulatory clearance, integration into clinical workflows, or liability, so the near-term effect is on research and tooling, not on patients being diagnosed differently tomorrow.
What is noise
The claim that Damo Radar "outperformed most radiologists" is asserted without a stated comparator group, reader study design or confidence intervals, so it cannot be verified from what's given. "World's first expert-level generalist medical imaging model" is a self-awarded superlative from the research team, and the idea that the training method will "extend to other imaging types" is aspiration, not demonstrated result. The secondary-source writeup also omits basic verification anchors: no link to the Science paper, the GitHub repo or the model weights.
Watch next
- 01Direct link to the Science paper and independent replication or critique of the 0.913 AUC figure and its comparator methodology
- 02Confirmation the model weights and code are actually accessible on a public repo (e.g. Hugging Face, GitHub) with a usable licence, not just announced
- 03Any hospital, health system or regulator (FDA, NMPA, EMA) engaging with Damo Radar for pilot use or clearance in the next 6-12 months
Coverage
1 storyMore capability signals
Full feed →- Deepseek releases V4.1-Flash, an open-source model that sharply cuts KV cache memory and input-processing compute for AI agents10 Sept 202682
- Anthropic threat report: Claude abused for malware, drone/missile software, mass surveillance, and industrial-scale distillation by Chinese AI labs11 Sept 202680
- WIRED investigation: Flock Safety's AI person-search tools let police run broad description-based surveillance, with weak guardrails against misuse3 Sept 202680
- Google DeepMind launches AlphaGenome Atlas, a free public database of predicted effects for 9 billion possible human genome variants8 Sept 202679