← Back to feed

Alibaba’s Damo Academy has open-sourced Damo Radar, a vision-language model for abdominal CT diagnosis that covers more conditions in a single pass than any prior radiology AI. The research appears in Science, the peer-reviewed journal, alongside a full performance evaluation across 40,000 real-world clinical exams.

What It Does

Damo Radar reads contrast-enhanced CT scans and identifies findings across 18 abdominal anatomical structures. The model handles 146 distinct clinical findings — including malignant tumours, pancreatic cancer, fatty liver disease, and acute appendicitis — using a single generalised architecture instead of a separate model per condition.

Key Numbers

• Mean AUC: 0.913 across 146 clinical findings (AUC of 1.0 = perfect diagnostic accuracy)
• 40,000 real-world examinations used for evaluation
• 23 of 26 radiologists outperformed in head-to-head comparison
• +10% detection sensitivity when radiologists use the model as a second read
• -30% diagnosis time in radiologist-plus-model workflows
• Trained on 400,000 contrast-enhanced CT exams paired with 15 million anatomy-aware image-text pairs

How It Was Built

The training method links medical images directly to clinical reports without manual per-finding annotation — a vision-language learning approach that turns raw CT volume and associated text into a generalised understanding of abdominal pathology. CT images are converted into 3D anatomical units, then precisely aligned with report text, which the team says is the core technical contribution that enabled scaling to 146 simultaneous findings.

Damo Academy calls it “the world’s first expert-level generalist medical imaging model.” That framing is vendor language, but the Science publication and the head-to-head radiologist data are independently significant.

Licensing and Access

The model weights and training code are on GitHub under Apache 2.0. The repository includes training, inference, and preprocessing guides, along with evaluation on the MERLIN external test set.

Prior radiology AI has been heavily disease-specific. A model trained to detect liver tumours typically does not detect pancreatic cysts with any reliability. Damo Radar’s 146-finding coverage across 18 structures is the practical gap it is designed to close. The team says the training framework can extend to other imaging modalities beyond abdominal CT — making generalisation to chest or brain scans a plausible future direction.

The immediate clinical question is prospective validation: AUC measured on retrospective exam sets does not always translate to real-time workflow performance. The radiologist time-reduction figure (30%) is the more operationally meaningful number if it holds in live deployment.