By Alibaba CloudLearning Objectives
- Understand what a generalist imaging model is and how it differs from single-disease tools
- Read the published results critically, including what a reader study does and does not show
- Check the license before planning research or a product around it
What Is DAMO RADAR?
DAMO RADAR is a vision-language model for abdominal CT from Alibaba's research institute, DAMO Academy. It was trained on more than 400,000 contrast-enhanced abdominal CT exams paired with the radiologists' written reports, which let it learn from ordinary clinical records rather than from images someone labeled by hand. From a single scan it flags 146 clinical findings across 18 organs, from cancers of the pancreas and liver to fatty liver disease and acute appendicitis.
Most medical imaging AI works one disease at a time: a tool for lung nodules, another for strokes, another for breast screening. RADAR's claim is that one model can cover the breadth of an abdominal read. DAMO Academy calls it the world's first expert-level generalist medical imaging model, and the paper appeared in the journal Science in September 2026, alongside the open-source release.
⚠️Warning
This is a research model, not a cleared clinical product. Neither the GitHub repository nor the Hugging Face model card claims any regulatory clearance, while the other tools in this category are commercial products built for clinical deployment. The published results describe performance in a study. They do not make RADAR suitable for reading patients' scans outside a research setting.
📝Note
The license is split, and the weights are non-commercial. The code on GitHub is Apache 2.0. The model weights on Hugging Face are released under Creative Commons BY-NC-SA 4.0, which forbids commercial use and requires anything built from them to carry the same terms. A hospital research group can study and adapt it; a company cannot ship it in a product under that license.
🎯Tip
Explore DAMO RADAR: the code and documentation are on GitHub and the model weights are on Hugging Face.
Pricing
There is no charge to download or run RADAR. What limits use is the license rather than a price.
- Training, inference and preprocessing code
- Apache 2.0 license
- Commercial use of the code allowed
- RADAR and RADAR+ checkpoints
- Creative Commons BY-NC-SA 4.0
- Non-commercial use only
Core Features
One Pass, Many Findings
A single scan returns results across 146 findings in 18 anatomical structures, instead of running a separate model for each condition.
Learns From Reports, Not Hand Labels
Training pairs images with the written clinical report, which avoids the expensive step of having specialists annotate every scan.
Organ-Level Alignment
DAMO Academy says the model converts CT images into three-dimensional anatomical units and aligns each one with the matching part of the report, which is how it keeps findings tied to the right organ.
A Public Benchmark Variant
The release includes RADAR+ checkpoints trained on, or fine-tuned for, the public Merlin CT dataset, so outside groups can compare against it on data they can obtain themselves.
Strengths
- Peer-reviewed — published in Science, rather than reported only by the developer
- Tested at scale — a mean area under the curve of 0.913 across 146 findings on nearly 40,000 real-world exams
- Compared with radiologists — it outperformed 23 of 26 radiologists in the study
- Helps people as well as competing with them — in the study, radiologists working with it raised detection sensitivity by 10 percent while taking 30 percent less time
- Open code and weights — anyone can download it and check the results
Limitations and Considerations
- Not cleared for clinical use — no regulatory clearance is claimed, so it is a research tool
- Non-commercial weights — the Creative Commons license rules out using the released model in a product
- One body region — it reads contrast-enhanced abdominal CT only; the authors say the approach could extend to other imaging, but that is a plan, not a result
- Study performance is not deployment performance — a single evaluation, however large, does not show how it behaves on another hospital's scanners, protocols and patients
- Area under the curve is an average — a strong mean across 146 findings can hide weaker results on individual conditions, so check the per-finding numbers for the ones you care about
Best Use Cases
| Use Case | Why DAMO RADAR Matters | Caveat |
|---|---|---|
| Radiology AI research | A strong open baseline across many findings | Non-commercial weights only |
| Benchmarking a commercial tool | RADAR+ runs on the public Merlin dataset | Match the scan protocol before comparing |
| Teaching medical imaging AI | Open code for training and inference | Not for reading real patients' scans |
| Studying report-supervised learning | Shows labels can come from clinical reports | Report quality limits what it learns |
Key Takeaways
- DAMO RADAR is Alibaba DAMO Academy's generalist model for abdominal CT, flagging 146 findings across 18 organs in a single pass
- It was published in Science in September 2026 and outperformed 23 of 26 radiologists in the study, with a mean area under the curve of 0.913 across nearly 40,000 exams
- Radiologists working with it raised detection sensitivity by 10 percent while taking 30 percent less time
- The code is Apache 2.0 but the weights carry a non-commercial Creative Commons license, so commercial use of the released model is not permitted
- It is a research model with no regulatory clearance, unlike the commercial imaging tools it sits beside
