RADAR: An Expert-Level Generalist AI for Abdominal CT Diagnosis

A generalist vision-language model for abdominal CT, published in Science on 2026-09-17 (vol 393, issue 6817, eaec6129) by Alibaba DAMO Academy with Zhejiang University's First Affiliated Hospital. RADAR is trained on more than 400,000 contrast-enhanced abdominal CT examinations and 15 million anatomy-wise image–text pairs, "learning directly from clinical reports without manual annotation" — the supervision comes from radiology reports rather than per-voxel labels, which is what lets one model cover the whole abdomen instead of a single task. Across internal and external evaluations at multiple centers the paper reports high diagnostic performance and generalization over 18 anatomical structures and 146 imaging findings, and in a reader study RADAR assistance raised the diagnostic sensitivity of 26 radiologists by about 10%. The authors' claim is that a generalist model "can match human experts in general and complicated radiology tasks."

Weights, code and preprocessing pipelines are public under CC BY-NC-SA 4.0 (research use only; the model card states that further improvement and prospective clinical studies are needed before clinical deployment). The release ships the RAD-CT-pretrained RADAR checkpoint, a UNet vision-branch checkpoint, and two RADAR+ checkpoints — one trained from scratch on Merlin-CT-Train and one pretrained on RAD-CT then fine-tuned on it — with BERT tokenizers and encoders for Chinese and English; the implementation builds on LAVIS, nnU-Net, MONAI and 3D-ResNets-PyTorch. Documentation covers training, inference against the external MERLIN test set, and report/image preprocessing. The paper has 40 authors: first author Qi Zhang and senior author Tingbo Liang are at the Department of Hepatobiliary and Pancreatic Surgery of Zhejiang University School of Medicine's First Affiliated Hospital, while the model work and senior author Ling Zhang are at Alibaba DAMO Academy (several DAMO co-authors hold joint Hupan Laboratory and Zhejiang University appointments). The GitHub repository, opened in July 2026, has about 149 stars.