Alibaba's RADAR matches 23 of 26 radiologists on abdominal CT
Medical imaging AI has been stuck for a decade: one disease, one model, a radiologist labeling every slice by hand. A paper in Science this week argues the generalist route is the one that actually scales.
Alibaba's Damo Academy has published a generalist radiology model in Science that reads abdominal CT scans about as well as most working radiologists — and open-sourced the weights, the code and the training framework. The model is RADAR, for Rapid Abdominal Diagnosis with AI and Radiology, and it covers 146 diseases across 18 anatomical structures in the abdomen. That region is the read radiologists openly dread: liver, gallbladder, pancreas, spleen, kidneys and intestine are packed together, coiled through the whole cavity, and almost entirely soft tissue of near-identical density. Xiao Wenbo, who runs radiology at the First Affiliated Hospital of Zhejiang University and has spent more than 30 years in the field, said junior doctors fear being assigned to the abdominal group more than any other.
The validation numbers are the reason this cleared Science's bar. On just under 39,000 consecutive real-world cases inside the network, RADAR posted a mean AUC of 0.913 — a measure where 0.5 is a coin flip and anything above 0.9 counts as excellent. On more than 24,000 CTs from eight outside hospitals, across different regions and different scanners, it held at 0.895. It also generalized to the emergency room, a setting it was never trained on: triage CTs there are rushed, the contrast agent often hasn't reached peak effect and patients may not be able to cooperate. On 27,000 emergency scans, AUC was still 0.904.
The sharpest test was the human comparison. Twenty-six radiologists from 14 hospitals read the same cases independently, and RADAR's accuracy came out ahead of 23 of them; it finished slightly behind three senior doctors. In a combined setting the arithmetic changed: with the model in the loop, sensitivity rose by roughly 10% and average reading time fell by more than 30%. The configuration the team flags — a junior doctor plus RADAR — beat a senior doctor reading alone. For a full abdominal CT report that takes a senior radiologist at least 20 minutes of working through hundreds of images, that is the number hospital administrators will remember.
The method is where the generalist claim actually lives. Older medical imaging AI is supervised learning: a doctor labels each scan, and the model can only recognize the diseases that were labeled. RADAR learns instead from the pairing that already exists in every hospital — the scan and the report written about it — with no extra annotation. The naive version of that fails, because aligning a whole CT series against a whole report drowns a small lesion in normal tissue. RADAR's fix is organ-level alignment: locate each organ first, split the series into organ units, then match each unit to the sentence describing it. It is modeled on how radiologists actually read — liver, then pancreas, then biliary tract, then kidney, then bowel, one at a time. A second piece, adaptive contrastive modeling, stops the model from treating two healthy livers as maximally different just because they belong to different patients.
Two caveats worth holding onto. The human-comparison result is reported by the authors, and beating 23 of 26 doctors in a retrospective read is not the same as outperforming them in live clinical practice. And the concern the Zhejiang team raises is not accuracy — it is training. If junior doctors lean on a model that writes at senior level, they can skip the years of pattern-building that produce the judgment to catch it when it is wrong. Xiao's answer is to use RADAR in reverse: write the report, then let the model mark what you missed, so every case doubles as a check instead of a shortcut.
The scaling curve in the paper is still climbing with no sign of saturation, and the framework is not locked to CT — the same organ-level alignment should transfer to MRI, PET and ultrasound wherever there is data to pair.
What to watch: whether a Chinese hospital network formally deploys this at scale, and whether the same organ-level alignment shows up in Western health systems that have far less report-and-scan data to train on.
Would you trust a generalist model to flag what your doctor missed — or does that shorten the training that makes a good doctor? Tell us in the comments.
Sources: RADAR (GitHub)
