Alibaba DAMO Academy and the First Affiliated Hospital of Zhejiang University published a paper on the DAMO RADAR model in the journal Science and simultaneously open-sourced its code. This is a vision-language model for analyzing contrast-enhanced abdominal CT scans: a single model covers 146 clinical findings across 18 organs, from malignant tumors to acute appendicitis, and in a blind comparison it outperformed most of the participating radiologists. The code is released under the Apache 2.0 license, checkpoints are available for download, but the model is not certified for clinical use.

image
image

What happened

The paper on DAMO RADAR was published in the journal Science: volume 393, issue 6817, DOI 10.1126/science.aec6129. The model was trained on more than 400,000 CT studies, which yielded 15 million "anatomy-text" pairs; training was conducted directly from clinical reports, without manual annotation. DAMO RADAR covers 146 clinical findings across 18 abdominal organs, including malignant tumors, fatty liver disease, and acute appendicitis. The average AUC was 0.913 across 146 findings on nearly 40,000 real examinations. In a blind comparison, the model outperformed 23 of 26 radiologists in average accuracy. The authors also report that with the model's help, doctors missed findings 10% less often and read scans 30% faster. The code is published in the alibaba-damo-academy/damo-radar repository on GitHub under Apache 2.0, and checkpoints are released on Hugging Face under the name radar-generalist.

Context

Radiology AI has so far been built as a set of narrow disease-specific models: a separate model for one finding, another for another, and such a zoo is expensive to maintain and scale. The second chronic bottleneck in medical imaging is annotation: high-quality image labels require doctors' time and are expensive. RADAR addresses both problems: the authors position it as the first "expert-level" generalist approach to medical imaging — this is their formulation, but it points to a real structural problem in the industry — and training directly from clinical reports eliminates the need for manual annotation. The authors state that the methodology is transferable to other types of medical imaging, meaning the significance of the work potentially extends beyond abdominal CT. Publication in Science with open code and weights is a rare case in this field: usually models of this level remain closed.

Why this matters for the industry

For the industry, the main signal is not in the headline "model outperformed radiologists," but in the combination of access and method. Code under Apache 2.0, checkpoints, and demo data give engineers a ready-made open-source baseline in abdominal CT: a prototype can be built today, without someone else's closed API, and the training method from clinical reports is claimed to be transferable to other modalities. The open release of a generalist model moves narrow radiology AI for a single finding into the commodity layer: value for companies shifts up the stack, into certification, integration into clinical workflow, and proprietary fine-tuning, rather than into the weights themselves. Narrow CT-AI vendors get a public point of comparison, and their negotiating position on price weakens. In the coming months, independent replications, comparative benchmarks, and the first fine-tuned models for local scanners and protocols are expected.

Why this matters for users

Checkpoints on Hugging Face and code on GitHub can already be downloaded, the claimed results can be reproduced on demo data, inference speed can be measured on your own hardware, and integration into your own pipeline can be assessed — a rare case where a model from a Science publication is actually available, rather than remaining behind a closed API. Based on it, a demo viewer or benchmark harness can already be built today. For patients and doctors in clinics, nothing changes: this is a research model, not a certified medical product, and it cannot be used in clinical practice now. But it makes sense to familiarize yourself with the methodology in advance: if the generalist approach is confirmed, such models will become the basis for future clinical products.

What is still unknown / limitations

Validation was conducted only on the authors' data: there are no independent replications and external checks on samples that do not overlap with the training set yet, so the average AUC of 0.913 still needs to be confirmed. Reproduction on demo data confirms the functionality of the artifacts, but not the claimed metrics on real examinations: "downloaded and ran" is not the same as "verified," clinical validation and technical verification are different levels of evidence. The formulation "first expert-level generalist approach" is the authors' positioning, not an independent assessment. The model is not certified as a medical device, its content is distributed under CC BY-NC-SA, and there is no data on API, price, and latency in the sources. Finally, the SCMP publication is owned by Alibaba, which should be taken into account when reading the primary source.

Sources

Author

Look at AI, editorial team