The takeaway: Researchers at Alibaba's DAMO Academy have published a single AI model that reads abdominal CT scans and looks for 146 different imaging findings across 18 anatomical structures; DAMO says these include cancers. The work appeared in Science on 17 September. The model weights have been released on Hugging Face under a non-commercial licence (CC BY-NC-SA 4.0), so they can be used for research and other non-commercial work, but not in commercial products without separate permission. The code on GitHub is under the more permissive Apache 2.0 licence.

What RADAR is

According to the paper's abstract, RADAR is a "generalist vision-language model" trained on more than 400,000 contrast-enhanced abdominal CT examinations and 15 million anatomy-wise image-text pairs. It learns "directly from clinical reports without manual annotation". In other words, it learns from the reports radiologists already write, rather than from scans that someone has labelled by hand for one disease at a time.

That is the pitch. DAMO says existing AI models typically target individual diseases, Yicai reports; Yicai writes that RADAR could potentially replace several disease-specific models with a single system.

What the paper reports

From the Science abstract:

• Across internal and external evaluations at multiple centres, RADAR "achieved high diagnostic performance and robust generalization" for 18 anatomical structures and 146 imaging findings.
• In a reader study, RADAR assistance increased the diagnostic sensitivity of 26 radiologists by about 10%.

Yicai's report, based on DAMO's announcement, adds more figures:

• A mean area under the curve (AUC) of 0.913 across the 146 findings, in nearly 40,000 real-world examinations. AUC measures how well a model separates positive from negative cases; 1 is perfect.
• Compared with the 26 radiologists, the model's average performance exceeded that of 23 of them.
• The study showed RADAR could help radiologists increase detection sensitivity by 10% while reducing the time needed by 30%.

We haven't been able to read the full paper, so these extra figures come from Yicai's reporting rather than from the paper directly.

How open is it?

The GitHub repository provides training code and an inference demo; the pre-trained checkpoints are on Hugging Face. The code on GitHub is released under the Apache 2.0 licence, which allows commercial use. (The repository's badge shows CC BY-NC-SA 4.0, but its LICENSE file is Apache 2.0.) The trained model weights on Hugging Face are under CC BY-NC-SA 4.0, which bars commercial use. DAMO and some coverage describe the release as open source. In our reading, "open code, with non-commercial weights" is the more precise description.

Why it matters

In our view, for patients the relevant question isn't the AUC. It's whether a tool like this catches things that would otherwise be missed, in real hospitals, without adding false alarms. In our reading, the reader-study result, radiologists becoming more sensitive when assisted, is the more clinically meaningful number of the two. The abstract doesn't state how assistance affected false positives, and we couldn't check that in the full paper.

In our reading, a model like this would also need to go through medical-device regulation before routine clinical use, in the UK or anywhere else. Nothing in the GitHub or Hugging Face pages we read says it has.