A generalist model for abdominal CT scans: RADAR's promises, between Science and licence terms

On 18 September 2026, the DAMO Academy of the Alibaba group released the code and weights of RADAR, a vision-language model designed to analyse contrast-enhanced abdominal CT scans. Unlike the usual practice of industry launches, the system's release was preceded by the publication of its results in the journal Science. According to the official documentation, the model was trained on more than 400,000 radiological examinations and 15 million image-text pairs, learning directly from clinical reports without manual annotation, in order to assess 18 anatomical structures and 146 radiological findings.

The figures published in the paper indicate a mean area under the ROC curve of 0.913 across the 146 findings in the internal cohort, with a 95% confidence interval between 0.911 and 0.915, and values ranging from 0.874 to 0.912 in tests carried out at eight external centres. For four types of tumour confirmed by pathology — liver, pancreas, stomach and colorectum — the figure runs from 0.891 to 0.984. In a reader study involving 26 radiologists, the model's standalone accuracy exceeded that of 23 of the 26 doctors taking part, while its assistance raised the group's diagnostic sensitivity by roughly 10 percentage points and cut reporting times by more than 30%. These are, however, aggregate metrics stated by the authors: independent verification on cohorts not selected by the research team is still missing. There is also a gap in the reported data volume: some of the press says 420,000 examinations, while the official repository and the Science paper say “more than 400,000”. The discrepancy cannot be settled against the paper's full text, which is not accessible to automated reading. The DAMO Academy team, quoted by the South China Morning Post, calls it “the world's first expert-level generalist medical imaging model”.

When it comes to actual adoption, methodological and legal constraints emerge. While the source code is distributed on GitHub under the Apache 2.0 licence, the weights published on Hugging Face are bound by the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 licence, which forbids commercial exploitation; the repository also points to a Zenodo archive of the material. It is not yet clear whether that restriction allows the technology to be used inside Europe's public healthcare facilities, which have to balance clinical innovation against the requirements of the medical devices regulation and the AI Act. On regulatory approvals for clinical use outside China we found no primary-source confirmation either way: the point remains open.

The sequence chosen — peer review first, release afterwards — is the reverse of how industry launches usually go. What remains to be seen is whether free access hedged by a non-commercial licence will be a resource for research labs or a legal puzzle for hospital wards.

— Olya