The model stayed accurate on emergency scans it was never trained on and at outside hospitals, the study in Science reports.
RADAR, from Alibaba's DAMO Academy, checks contrast-enhanced CT scans of the abdomen for 146 kinds of findings, malignant tumors among them. Its authors note that most radiology AI is trained for one narrow task. Their study appeared in Science on 17 September, and the code is on GitHub.
The model learned from 424,911 abdominal CT examinations and the reports doctors had already written about them, split into 15 million image-and-text pairs, each tied to one of the 18 organs and other structures it covers. Nobody labeled the scans by hand.
According to a summary from AAAS, which publishes Science, RADAR scored a mean AUC of 0.913 across the 146 findings, against 0.776 for the systems it was compared with. AUC measures how well a test tells positive cases from negative ones, where 1 is perfect and 0.5 is chance. The model reached 0.904 on more than 27,000 emergency scans, a setting it had no specific training for, and 0.895 at hospitals outside the study group.
In the authors' reader study, 26 radiologists working with RADAR's suggestions raised their sensitivity, the share of real findings they caught, by about 10%.
The code carries the Apache 2.0 license. The trained weights on Hugging Face do not: they are released under CC BY-NC-SA 4.0, which covers research use and leaves commercial deployment to a separate agreement.