Science Publishes DAMO RADAR, a General Medical Imaging AI Model from Alibaba DAMO Academy

On September 18, 2026, Science published research on DAMO RADAR, a general-purpose medical imaging AI model developed by Alibaba DAMO Academy in collaboration with the First Affiliated Hospital of Zhejiang University School of Medicine and other institutions. Described as the first expert-level general medical imaging AI model, it covers 18 abdominal anatomical structures and 146 diseases. The model, code, and framework are fully open-sourced.

Abdominal CT is considered one of the most challenging imaging scenarios, as multiple systems are involved and organs are mostly soft tissues with similar densities. Instead of relying on supervised learning with labor-intensive annotations, RADAR uses vision-language contrastive learning to train on CT images paired with diagnostic reports. Its core design involves organ-level fine-grained alignment, which splits a CT volume into organ units and aligns each with corresponding text descriptions, mimicking how radiologists read scans organ by organ. The model also uses adaptive contrastive modeling to adjust sample distances based on medical knowledge.

In validation, RADAR achieved an average AUC of 0.913 in an internal cohort of nearly 39,000 consecutive cases and 0.895 on more than 24,000 CT scans from eight external hospitals. It also reached an AUC of 0.904 on 27,000 emergency cases not included in training. In a harder pathological gold-standard evaluation on four cancer types, AUC ranged from 0.891 to 0.984.

In a human-machine comparison involving 26 radiologists from 14 hospitals, RADAR outperformed 23 of them and trailed only three senior doctors. When used in a human-AI collaborative setting, radiologists' overall sensitivity increased by about 10% and average reading time dropped by more than 30%. Xiao Wenbo, director of radiology at the First Affiliated Hospital of Zhejiang University School of Medicine, said the AI tool can help reduce missed lesions and support junior radiologist training.

The team said the organ-level alignment framework could also be adapted to other imaging modalities such as MRI, PET, and ultrasound. DAMO RADAR's model, code, and technical framework are all available as open source.