Alibaba's RADAR AI hits expert level on abdominal CT

RADAR is presented as a generalist vision-language model, meaning one system handles a range of reading tasks rather than a single narrow one. The scale of its training set — over 400,000 contrast-enhanced abdominal CT examinations and 15 million anatomy-aware image–text pairs — is what the release points to as the basis for its breadth, and the model learns directly from clinical reports rather than scans that humans have labeled by hand. That matters because hand-labeling is the slow, expensive step that keeps most radiology AI tools confined to one task.

The release ships with a setup path, HuggingFace checkpoints, helper download scripts, and documentation covering training, inference, and preprocessing. RADAR can be trained from scratch, fine-tuned on MERLIN data, or run for inference and evaluation on the external MERLIN test set, which is likely to be the first real checkpoint on whether the expert-level claims travel beyond the data the model was built on.

For radiologists and radiology AI groups, the practical appeal is access: an Apache-licensed model with reusable preprocessing code lowers the cost of testing whether a generalist reader helps in their own clinical setting. The open question hangs on external validation and on how the model would be integrated into reporting workflows, neither of which the release resolves on its own.