Alibaba open-sources DAMO Radar: the AI that reads abdominal scans and beats 23 out of 26 radiologists, study published in Science
🔎 An expert medical AI, peer-validated, and free
The timing is almost ironic. On September 20, 2026, the Financial Times reported that clinicians are holding back the expansion of medical AI beyond diagnosis and imaging, citing a lack of clinical data and peer-verifiable performance. Two days earlier, Alibaba's DAMO Academy had just published in Science — meaning peer-reviewed — exactly the type of evidence they are calling for.
This model, named RADAR (Rapid Abdominal Diagnosis with AI and Radiology), reads contrast-enhanced abdominal CT scans covering 18 organs and identifies nearly 150 conditions, including cancers. Trained on more than 420,000 real examinations, it posts an average AUC of 0.913 across 146 clinical findings and outperformed 23 of the 26 expert radiologists in comparative reading, according to TechTimes.
And the detail that changes everything: Alibaba has open-sourced the weights, the code, and the training framework. The model is available at no cost to hospitals and researchers worldwide. In a healthcare AI market estimated at $36.67 billion in 2026 (MarketsandMarkets, via Cryptopolitan), it's a bombshell.
Key Takeaways
- A generalist medical VLM: RADAR analyzes contrast-enhanced abdominal CT scans across 18 organs and identifies 146 clinical findings, including malignant tumors.
- Expert-level numbers: average AUC of 0.913 across 146 findings (vs. 0.776 for the best competing VLM), and a win over 23 of the 26 specialist radiologists in comparative reading.
- Validated in Science: study published and announced on September 18, 2026, tested on roughly 40,000 real-world exams, with internal and external multi-center evaluations.
- 100% open source: weights, code, and training framework published on GitHub and Hugging Face, free for hospitals and researchers.
- Acknowledged limitations: contrast-enhanced abdominal CT only, no FDA clearance, local computing infrastructure required for volumetric inference.
Recommended Tools
| Resource | Main use | Price (September 2026) | Ideal for |
|---|---|---|---|
| DAMO Radar on GitHub | Weights, code, and training framework | Free (open source, check github.com) | Researchers, hospitals with GPU infrastructure |
| Science Abstract via PubMed | Full methodology and evaluations | Free | Verify the numbers at the source |
| EurekAlert Press Release | Dataset and benchmark details | Free | Understand the organ-level alignment |
| AI Weekly | Daily AI and healthcare news digest | Free | Track upcoming deployments |
What exactly is DAMO Radar?
It's a generalist medical vision-language model (VLM), trained by Alibaba's DAMO Academy to read abdominal CT scans and produce clinical conclusions covering 18 organs and nearly 150 conditions — from malignant tumors to more mundane anomalies.
Where most medical AIs excel at a single task (detecting a nodule, measuring a lesion), RADAR lays claim to generality. The full name published on PubMed — RADAR, for Rapid Abdominal Diagnosis with AI and Radiology — embraces the ambition: a single system for the entire abdomen.
The most interesting technical detail: the model learned directly from radiology reports, without manual annotation. No teams of clinicians paid for years to label images. Existing reports serve as supervision, which explains the scale of the dataset: more than 420,000 exams and more than 15 million image-text pairs at the anatomical level, according to the official press release.
This approach solves two problems at once. First, the annotation bottleneck: no need to create labeled datasets — you leverage the ones that already exist. Second, signal quality: a radiology report encodes an expert's reasoning, not just a binary answer.
The trick that changes everything: organ-level alignment
Why do existing generalist VLMs fail at radiology? Because a CT scan isn't an image — it's a sparse 3D volume: hundreds of slices, organs of radically different sizes and densities. Models designed for flat 2D images drown in it.
Qi Zhang's team worked around the problem in two steps, as detailed by Pandaily. First, each 3D volume is decomposed into individual anatomical structures. Then, images and report texts are paired at the organ level via large-scale contrastive learning, refined by adaptive contrast modeling.
The result: the model doesn't "see" an indistinct abdomen — it sees 18 organs, each tied to the clinical vocabulary that describes it. It's this fine-grained alignment that would explain the performance gap with competing VLMs — 0.913 versus 0.776 average AUC.
The numbers that matter — and those you shouldn't confuse
RADAR underwent two distinct evaluations, and conflating them is the primary source of misinterpretation about this model.
The first is a multi-center benchmark, internal and external, covering 146 abdominal CT reports. RADAR achieves an average AUC of 0.913, versus 0.776 for the best competing VLM, according to EurekAlert. AUC measures the ability to distinguish positive from negative cases; 0.913 on average across 146 heterogeneous conditions is a very high-level score.
The second is a comparative reader study, published in Science: in head-to-head comparisons on the same cases, RADAR outperformed 23 of 26 specialist radiologists, reports TechTimes. This is the number that makes the headlines — and it comes from a study that is separate from the benchmark against other AIs.
Finally, the full-scale test: roughly 40,000 real-world examinations, according to the South China Morning Post. Not a sanitized academic benchmark, but hospital scanners, with their noise and edge cases.
| Evaluation | Result | What it measures |
|---|---|---|
| Multi-center VLM benchmark (146 reports) | Average AUC 0.913 vs 0.776 for the best competitor | Diagnostic discrimination vs other AIs |
| Comparative reader study (Science) | RADAR ahead of 23 of 26 expert radiologists | Performance vs human specialists |
| Real-world conditions | ~40,000 real-world examinations | Robustness outside the lab |
My reading advice: always cite the evaluation along with the number. "0.913" without "average AUC across 146 reports" means nothing, and neither does "23 out of 26" without "reader study published in Science".
Why open source is the real masterstroke
Because the weights, code, and training framework are published, RADAR is not a demo — it's a research infrastructure that any hospital or lab can audit, reproduce, and improve.
In closed medical AI, the usual business model is a per-exam subscription or a per-institution license. Here, the marginal cost comes down to your own computing infrastructure. For budget-constrained public hospitals — the majority of hospitals on the planet — the difference isn't marginal, it's existential.
Auditability counts double in medicine. A model whose weights are public can be tested on an institution's local data, on its own patient populations, before any clinical decision. This is precisely the kind of local validation that emerging regulatory frameworks are starting to require.
This release is part of a broader wave. Open source is no longer just about code models like Qwen3 Coder Next, which runs on a single 64 GB Mac or Kimi K2.7-Code with its trillion parameters, nor research agents like ByteDance's DeerFlow. It now reaches the most regulated domain there is: healthcare.
My take: this is Alibaba's most strategic decision in this whole story. By open-sourcing an expert-level model, DAMO Academy is buying what no marketing budget can buy — the trust of the scientific community. AI Weekly called it the standout story of the week, and deservedly so.
The context: AI medicine has a proof problem
RADAR arrives at the exact moment when the clinical literature is beginning to document the gap between the accuracy of medical AIs and the proof of better patient outcomes.
The Financial Times of September 20, 2026 summed it up: clinicians express significant reservations about the expansion of AI beyond diagnosis and imaging, for lack of sufficient clinical data to verify the real-world effectiveness of these technologies.
The hard numbers, compiled by Cryptopolitan, speak for themselves. The Penda Health trial in Kenya, published in Nature Medicine on June 26, 2026, covering 9,691 patients: a 14-day treatment failure rate of 2.2% with AI versus 2% without — a non-significant difference (aOR 0.77). The 2024 RAPIDx AI trial with 3,029 patients: 26% versus 26.4% composite outcomes at 6 months, even though invasive angiography was 47% less frequent with AI.
In other words: AI diagnoses well, but improving diagnosis alone isn't enough to improve the patient's fate. That's the crux of the problem.
The study by Li Zhang, Jakob Kather et al., also in Nature Medicine, sketches a way forward: 90.04% accuracy on a seven-disease benchmark, with an on-site agent that keeps only 49.4% of cases — but on those filtered cases, accuracy climbs to 98.9%. Triaging cases beats the pretense of handling them all.
The market, meanwhile, isn't waiting for the evidence: from $36.67 billion in 2026 to $194.79 billion in 2031, i.e., 39.7% compound annual growth according to MarketsandMarkets. And research is getting organized: Stanford had 37,000 AI agents found a virtual biotech to predict the success of clinical trials — because predicting a trial's outcome before launching it is becoming a major economic stake.
On the regulatory side, the British review published on September 10, 2026 by the National Commission into the Regulation of AI in Healthcare proposes a pragmatic model: provisional approval of innovative systems (controlled real-world use before full certification), continuous post-deployment monitoring, and greater transparency about safety incidents.
In this landscape, RADAR stands out as a welcome exception: an imaging model — the one domain where clinicians already accept AI — with a publication in Science, multi-center evaluations, and open weights. The peer-reviewed proof the FT was calling for, delivered turnkey.
What RADAR Can't (Yet) Do
Three limitations, acknowledged by the team itself, should temper the triumphant headlines.
A single type of scan. RADAR only works on abdominal CT scans with contrast agent. No MRI, no contrast-free CT, no imaging outside the abdomen. A patient for whom contrast is contraindicated simply falls outside the model's scope.
Tokenizers with a linguistic false bottom. The model embeds Chinese BERT tokenizers in addition to the English versions, a reflection of the radiology reports used in training. A French-speaking hospital will have to assess the reliability of the conclusions in its own linguistic context before any serious use.
A demanding local infrastructure. Volumetric inference requires local computing infrastructure, TechTimes reminds us. Unlike an API you call from a browser, RADAR requires on-site GPUs — an investment not every institution can afford.
Add the lack of FDA clearance in the United States, and the picture is clear: RADAR is a tool for research and controlled deployment, not a certified medical product ready for use.
As for the "human in the loop," the evidence map by Joy Xu, Justin Ko, and Joseph Kvedar reminds us via Cryptopolitan: it isn't enough if the clinician isn't equipped to question and stop the system. Without training and without a real power to halt it, the human in the loop is merely a "moral crush zone" — the moral crumple zone that absorbs responsibility when the machine gets it wrong.
What's next: chest CT, brain MRI, and the race for scientific AI
According to Zhang Ling, senior algorithm expert on the project cited by TechTimes, organ-level alignment is not limited to the abdomen or to CT.
The targeted extensions: chest CT, brain MRI, retinal fundus photography. If the method delivers on its promises beyond its original scope, RADAR won't remain a single model but will become a family — and the Science paper the methodological template for generalist radiological AI.
This movement is part of a broader acceleration in scientific AI. DeepMind has precomputed the molecular effect of 9 billion genetic variants with its AlphaGenome atlas, and Anthropic has quietly set up a physical biology lab. Tech giants are no longer content with selling AI to laboratories — they're doing the science themselves.
My bet: in eighteen months, organ-level alignment will be cited in a good share of general radiological AI papers. The question isn't whether the method will spread, but who will apply it first to the most strategic organs — and under what regulatory regime.
❌ Common Mistakes
Mistake 1: Confusing mean AUC with diagnostic accuracy rate
The 0.913 is a mean AUC across 146 findings — a measure of statistical discrimination, not a "91.3% success rate." An AUC of 0.913 on a common condition and on a rare disease do not carry the same clinical significance. The solution: always cite "mean AUC across 146 findings, multi-center evaluations" and link to the paper.
Mistake 2: Conflating the two studies
"RADAR beats 23 out of 26 radiologists" and "AUC 0.913 vs 0.776" come from two different evaluations: the first is a comparative reader study published in Science, the second a benchmark against other VLMs. Merging them ("it beats radiologists AND AI with 0.913") is factually wrong. Two numbers, two sources, two sentences.
Mistake 3: Deploying without a regulatory framework
"Expert-level" does not mean "certified." RADAR has no FDA clearance, and the UK review specifically recommends provisional approvals with continuous monitoring before any full certification. Deploying such a model for clinical decision-making without an equivalent framework exposes both the institution and its patients. The solution: research use and assisted reading first, validation on local data next.
Mistake 4: Believing that "human in the loop" is enough
Parking a radiologist in front of the screen solves nothing if they have neither the means to question the model nor the power to stop it. The evidence map by Xu, Ko, and Kvedar is unequivocal: the clinician must be equipped to challenge the system, otherwise they become an accountability sink. The solution: escalation protocols, rejection thresholds, and traceability of human-machine disagreements.
❓ Frequently Asked Questions
Does DAMO Radar replace radiologists?
No. It outperformed 23 of 26 specialists in a reading study published in Science, but only on contrast-enhanced abdominal CTs, without FDA clearance, and on a scope of 146 findings. The realistic scenario is assisted reading: the model pre-filters, the radiologist confirms or contradicts. The profession shifts, it doesn't disappear.
Is the model really free?
Yes. The weights, code, and training framework are published open source on GitHub and Hugging Face, and the model is immediately available at no cost to hospitals and researchers worldwide. The free part stops at the infrastructure: volumetric inference requires local GPUs, whose cost remains on you.
Can RADAR detect cancer?
Yes. Among the nearly 150 abdominal conditions covered across 18 organs are malignant tumors. The South China Morning Post notes that the model was tested on nearly 40,000 real-world scans and outperformed most radiologists according to the study published in Science. Limitation: detection on contrast-enhanced CT only.
What languages does it work in?
The model embeds Chinese BERT tokenizers in addition to the English versions, reflecting the radiology reports used in training. For a French-speaking hospital, the quality of findings in a different linguistic context remains an open question, to be evaluated on local data before any serious clinical use.
When will we see RADAR on other organs?
Zhang Ling indicates that organ-level alignment is not limited to the abdomen or to CT: the targeted extensions are chest CT, brain MRI, and retinal fundus photography. No timeline has been published, but since the method is open source, other teams can reproduce it right now.
How does it differ from closed medical AIs?
Three differences: auditability (public weights, testable on each hospital's local data), cost (free versus per-scan or per-facility licensing), and proof (publication in Science with multi-center evaluations). Closed vendors retain the advantage on turnkey clinical integration and regulatory support — for now.
✅ Conclusion
By publishing in Science an expert-level medical model, tested on 40,000 real exams and open-sourced with weights included, Alibaba's DAMO Academy is giving clinicians exactly the peer-reviewed proof they had been calling for — it's now up to you, hospitals and researchers, to turn it into responsible deployments. Download the weights on GitHub and follow the rest via AI Weekly.