Alibaba DAMO publishes RADAR abdominal CT model in Science

Alibaba’s DAMO Academy and hospital partners published RADAR (Rapid Abdominal Diagnosis with AI and Radiology), a generalist vision-language model for contrast-enhanced abdominal CT, in Science (Sep 17). The product event is a medical imaging model that covers 18 anatomical structures and 146 findings in one system, trained from clinical reports rather than hand-labeled single-disease datasets.

What it does: Trained on 424,911 contrast-enhanced abdominal CT exams with millions of image–text / anatomy-specific pairs. AAAS SciPak: mean AUC 0.913 across the 146 findings; AUC 0.895 at eight external centers; emergency-setting AUC 0.904 on >27,000 cases without emergency-specific training. Best competing vision-language model cited at 0.776 mean AUC. Reader-study framing in coverage: assistance lifted radiologist sensitivity ~10% and cut read time ~30%.

What shipped for people who touch it: The code is now public on GitHub at alibaba-damo-academy/damo-radar, and the model checkpoints are on Hugging Face and ModelScope. The repo and the Hugging Face model card list the license as CC BY-NC-SA 4.0, which bars commercial use. Both say the model “is currently intended for research purposes only” and needs prospective clinical studies before clinical deployment. This is a research release for hospitals and labs, not a consumer ChatGPT-style app.

Limits: Abdominal contrast CT only in this paper. Not a guarantee of bedside deployment, regulatory clearance, or commercial reuse. Model-vs-radiologist headlines need the study’s exact protocol.

Who it is for: Radiology researchers and health systems evaluating generalist CT assistance — not patients self-diagnosing.

Risk: Diagnostic AI that “beats radiologists” is a liability and workflow product. Missed findings, domain shift across scanners, and unlicensed commercial use are the failure modes. Treat AUC tables as study results under authors’ conditions.

What’s next: Any FDA/NMPA path, prospective clinical studies, and independent external replications.

Source: AAAS / EurekAlert SciPak, Sep 17, 2026; Science peer-reviewed package (Zhang et al.); GitHub repo; Hugging Face model card.

— Rowan Keel