Alibaba's DAMO RADAR reaches Science with a general abdominal imaging model, fully open-sourced

Science has published a study from China describing DAMO RADAR — short for Rapid Abdominal Diagnosis with AI and Radiology — which its developers call the first expert-level general medical imaging model. A single model covers 146 diseases across 18 anatomical structures in the abdomen, spanning the liver, pancreas, gallbladder, kidneys, spleen, and intestines. It was built by Alibaba's DAMO Academy together with Zhejiang University's First Affiliated Hospital and other hospitals, and the model, code, and technical framework have all been open-sourced.

The abdomen is widely regarded as the hardest imaging scenario in clinical practice. The digestive, urinary, and reproductive systems crowd together, the intestines coil through the whole cavity, and almost every organ is soft tissue of similar density, so readers rely on a fine sensitivity to density differences to spot lesions. Dedicated AI tools for lung nodules have already shipped, but the abdomen stayed a no-go zone. Zhang Jianpeng, a senior algorithm expert at DAMO Academy, said a single disease type can take two to three years to master, while a real CT scan often contains multiple organs and multiple lesions at once.

Methodologically, earlier medical imaging AI relied on supervised learning in which doctors annotated CT images slice by slice — expensive, and limited to the diseases that were annotated. RADAR instead uses vision-language contrastive learning, drawing directly on the large numbers of CT images and matching diagnostic reports that hospitals already generate. Its central advance is fine-grained organ-level alignment: the model first localizes organs such as the liver, pancreas, gallbladder, and kidneys, then aligns each organ unit with the corresponding description in the report. A second innovation, adaptive contrastive modeling, adjusts the distance between samples according to medical knowledge, so that two patients with healthy livers are not treated as unrelated.

Validation came in several layers. On an internal real-world consecutive cohort of nearly 39,000 cases, RADAR reached a mean AUC of 0.913. On more than 24,000 CT scans from eight external hospitals, covering different regions and scanner hardware, AUC still reached 0.895. Emergency scans were not part of the training objective, yet AUC on 27,000 emergency cases was 0.904. For liver, pancreatic, gastric, and colorectal cancers, measured against pathology as the gold standard, AUC ranged from 0.891 to 0.984.

The human comparison is striking. Twenty-six radiologists from 14 hospitals read the same cases independently alongside RADAR, and the model's accuracy exceeded that of 23 of them, falling slightly short of only three senior physicians. In a human-plus-AI setting, readers' overall sensitivity rose by roughly 10% and average reading time fell by more than 30%. A junior doctor paired with the model even surpassed a senior doctor working alone.

Clinicians on the ground give the study its most vivid footnote. Xiao Wenbo, director of radiology at Zhejiang University's First Affiliated Hospital, said she could not believe the validation data at first sight, and that many doctors in her department pushed to deploy the model immediately. She also flagged a subtler worry: not accuracy — a human still signs off — but the risk that junior doctors become dependent on the tool and lose the imaging reasoning they normally build over years. Her proposal is to use RADAR as a one-on-one teaching aid, letting trainees self-check their reports and see instantly which lesion they missed.

The work has not stayed behind a paywall. DAMO Academy has open-sourced the model, the code, and the framework, so other teams can reproduce, improve, and adapt it. The team says the organ-level alignment approach should transfer to other modalities, with MRI, PET, and ultrasound as natural targets given suitable data. Scaling curves in the paper show performance still rising with more data, with no sign of saturation.

Zhang Ling, a senior algorithm expert at DAMO Academy, noted that Science rarely publishes medical imaging AI papers because the field has long been treated as an engineering problem; RADAR, he said, is the first demonstration that general medical imaging AI is a realistic technical path, and it changed reviewers' minds about treating it as a science question. Medicine is the domain with the most extreme demand for precision and the lowest tolerance for error, so if a general-model paradigm works here, it is likely to carry over to other high-stakes settings. What to watch next: how many hospitals actually deploy the open framework, and how well it transfers to MRI, PET, and ultrasound.