TL;DR
Alibaba's DAMO Academy has open-sourced RADAR, a radiology AI that screens 146 diseases across 18 organs in a single abdominal CT scan, outperforming 23 of 26 specialist radiologists while cutting reading time by more than 30%.
What happened
- DAMO RADAR, published in Science on September 18, 2026, screens one contrast-enhanced abdominal CT for 146 disease categories across 18 organs.
- Trained on 420,000 CT examinations converted into roughly 15 million anatomy-aware image-text pairs, using existing hospital radiology reports rather than new manual labels.
- In a formal reader study, RADAR's diagnostic accuracy exceeded that of 23 out of 26 specialist radiologists.
- Across eight independent clinical centres and ~40,000 real-world exams, the model achieved a mean AUC of 0.913.
- Code is released on GitHub and Hugging Face under Apache 2.0 (code) and CC BY-NC-SA 4.0 (weights, non-commercial research use).
Why it matters
- The human-plus-AI result is the headline: radiologists using RADAR as a second reader gained roughly 10 percentage points in detection sensitivity while cutting read time by more than 30%.
- Junior radiologists with RADAR matched the accuracy of senior colleagues working without it, compressing the experience gap that drives diagnostic inequality.
- Most radiology AI is single-task; RADAR demonstrates that one model can cover a broad disease surface within a defined imaging type, a meaningful architectural shift.
- Open weights lower the barrier for hospitals and researchers to audit and adapt the system rather than accept vendor benchmarks at face value.
- Critical limits apply: validated only on contrast-enhanced abdominal CT, trained predominantly on Chinese (Zhejiang Province) data, no FDA clearance, and no prospective clinical-trial outcomes published yet.
What to watch next
- External validation in non-Asian patient populations: performance figures cannot be assumed to transfer until tested in diverse cohorts outside China.
- FDA and international regulatory submissions: the gap between a Science publication and clinical deployment in the US or EU is wide, and the timeline will signal how seriously Alibaba pursues commercial rollout.
- Community fine-tuning via the open release: research groups adopting the Apache 2.0 code could accelerate validation across new populations and imaging protocols, or expose failure modes the original study did not surface.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.