§ Claim under review · Release
"Alibaba's Damo Academy has open-sourced Damo Radar, an AI model that can detect nearly 150 abdominal medical conditions, including cancers, from CT scans"
Verdict
Mostly accurate
Confidence
HighSummary
This one largely checks out. Alibaba's DAMO Academy did publicly release a model called DAMO RADAR on September 18 2026, with the code on its official GitHub account under an open Apache licence and the model files on Hugging Face, alongside a paper in the journal Science. The model covers 146 abdominal findings across 18 organs and structures, cancers included, so "nearly 150 conditions" is fair. The reported average AUC of 0.913 and the result that it outperformed 23 of 26 radiologists are both figures the researchers published. Three things the post leaves out matter: the model is built for contrast-enhanced abdominal CT specifically rather than CT scans in general, the 0.913 score comes from the hospital where it was developed while performance at eight outside hospitals averaged a lower 0.895, and the claim that doctors cut missed diagnoses by 10 percent restates a roughly 10 percent gain in sensitivity, which is not the same measurement. Most importantly, this is a research release with no regulatory clearance found in any country and no published trial of its use in real clinical workflow, so it is not an approved diagnostic tool. One detail could not be confirmed: some reports say the model weights carry a non-commercial licence, which would make the release less fully open than the word "open-sourced" suggests.
The readings
key figures from the evidencemean AUC across 146 findings, internal cohort of ~39,000 cases
average AUC across eight external validation centers
Why this verdict
Evidence
The release is real and the paper is real. The official Alibaba DAMO Academy GitHub organization lists damo-radar as a public Python repository under an Apache-2.0 licence, last pushed September 18 2026, and repository activity from that date is visible. A Hugging Face organization named radar-generalist hosts the model, a README and an auxiliary dataset.
On what the model does, the vision-language model, called Damo Radar, was designed to analyse contrast-enhanced CT scans covering 18 abdominal organs and identify a broad range of diseases and other abnormalities, such as malignant tumours, according to the institute . The publisher's own summary states that Qi Zhang and colleagues present Rapid Abdominal Diagnosis with AI and Radiology (RADAR), a vision-language AI model designed to provide broad diagnostic interpretation of contrast-enhanced abdominal CT scans, trained on a dataset of 424,911 examinations containing 1.5 million image-text pairs and more than 15 million anatomy-specific pairs .
On the headline metric, the authors report a mean AUC of 0.913 across 146 abdominal CT findings, compared with 0.776 for the best competing vision-language model, and an AUC of 0.904 across more than 27,000 emergency CT cases despite not being specifically trained on emergency data . Separately, in testing in cohorts at eight external centers RADAR maintained an AUC of 0.895 . One detailed trade account places the 0.913 figure specifically in the development institution's own cohort: the paper reports an AUC of 0.913 in an internal consecutive cohort of nearly 39,000 cases, with validation across eight external medical centers covering over 24,000 cases , and across those eight external centers the average AUC was 0.895, with individual-center results ranging from 0.874 to 0.912 .
On the reader study, the model's average accuracy exceeded that of 23 of the 26 radiologists compared, and with AI assistance the radiologists improved sensitivity by about 10 percent, meaning they missed fewer positive findings, while cutting reading time by more than 30 percent .
On the release terms, the GitHub code licence is Apache-2.0 on the official organization page. Two low-profile trade sites report a different licence for the weights themselves: the pretrained weights are described as distributed under a CC BY-NC-SA 4.0 licence, which permits research and other non-commercial uses but places restrictions on commercial deployment . I could not retrieve the licence file to confirm this.
On clinical status, the model does not have US FDA clearance for clinical deployment, and prospective real-world outcome data from a registered clinical trial had not yet been published . The registered study record describes a retrospective multicenter cohort comprising approximately 2 million cases of multiphase contrast-enhanced abdominal CT, all paired with corresponding radiology reports , with inclusion limited to multiphase contrast-enhanced abdominal CT covering the full abdominal region with matched radiology reports .
Findings
✓ What's accurate 10
- Alibaba's DAMO Academy did publicly release DAMO RADAR. The code is on the official DAMO Academy GitHub organization under Apache-2.0, dated September 18 2026, and checkpoints plus an auxiliary dataset are on the official Hugging Face organization. This is the strongest element of the claim and it is verified on the vendor's own channels.
- The model is a vision-language model for abdominal CT that covers 146 findings, so "nearly 150" is a fair rounding.
- Cancers are within scope. The design target explicitly includes malignant tumours.
- The figure of 18 abdominal structures is supported.
- Training on CT scans paired with clinical reports is supported, at a scale of 424,911 examinations and over 15 million anatomy-specific image-text pairs.
- Mean AUC 0.913 across 146 findings is the figure the authors report.
- The reader study result, better average performance than 23 of 26 radiologists, is reported as stated.
- The roughly 30 percent reduction in reading time with AI assistance is reported as stated.
- Publication in Science is confirmed by the PubMed record and the journal's own editorial summary.
- The statement that the approach could extend to other imaging modalities is attributed to the researchers, and the caption correctly frames it as a forward-looking claim rather than a result.
≈ What's misleading 7
- **Omitted qualifier:** the claim says "from CT scans." The model is built and validated for contrast-enhanced abdominal CT specifically, and the registered study excluded scans outside that protocol. A reader would reasonably infer it works on CT scans generally, including non-contrast studies, which is not what was tested.
- **Omitted qualifier:** the caption attaches the 0.913 figure to "nearly 40,000 real-world exams" without noting that this is the internal cohort at the development institution, and that performance at eight external centers was lower, averaging 0.895 with a per-center floor of about 0.874. The external number is the one that speaks to use elsewhere, and it is the one omitted.
- **Capability extrapolation:** the caption glosses AUC 0.913 as meaning the model "was highly effective at distinguishing abnormal scans from normal ones." AUC is a per-finding threshold-independent discrimination measure averaged across 146 findings. It is not a scan-level normal versus abnormal accuracy, and it is not a percentage of patients diagnosed correctly.
- **Omitted qualifier:** the caption says doctors "reduced missed diagnoses by 10%." The reported result is an improvement in radiologist sensitivity of about 10 percent. A ten point sensitivity gain and a ten percent reduction in the number of misses are different quantities, and the caption picks the framing that sounds like the larger clinical effect.
- **Demo to product conflation, partial:** "an AI model that can detect nearly 150 abdominal medical conditions" reads as a deployable diagnostic capability. The release is a research artifact with no regulatory clearance identified anywhere, and no published prospective trial of use in routine workflow. The post does not claim clinical approval, but it also gives a reader no signal that this is not an approved diagnostic tool.
- **Definitional, on "open-sourced":** the code is genuinely Apache-2.0, which is an OSI-recognised open licence. Two trade sites report that the weights carry a CC BY-NC-SA 4.0 non-commercial licence, which would not be open source in the OSI sense and would bar commercial deployment without a separate agreement. If that report is correct, "open-sourced" is accurate for the code and loose for the model itself. I could not verify the weights licence at source, so this is flagged rather than asserted.
- A separate framing point on the underlying findings: the 146 items are radiological findings, a category that spans cancers, other diseases and non-disease abnormalities. Rendering all 146 as "medical conditions" slightly inflates the disease coverage, although it does not change the substance of the claim.
? What's uncertain 6
- The licence on the model weights. Two low-profile secondary sites report CC BY-NC-SA 4.0. I retrieved the Hugging Face organization page but not the licence file, so I record the Apache-2.0 code licence as verified and the weights licence as reported and unconfirmed.
- The full Science article is paywalled and I did not retrieve it. Every numeric detail above therefore comes from the journal's editorial summary, the AAAS press release, or trade reporting of the paper, not from the article text. The internal versus external cohort breakdown in particular rests on a single detailed trade account.
- The exact composition of the "nearly 40,000" test set, and whether the reader study cases were drawn from internal or external cohorts. Not published in any source I could reach.
- Whether the reported 10 percent sensitivity gain is expressed in percentage points or as a relative change. The available wording does not resolve this.
- Independent verification. No third party has yet published an evaluation of the released weights on an independent cohort, which is the evidence that would move the performance figures from author-reported to independently established.
- One commentary site misidentified the underlying paper as a liver-focused Nature Medicine study. That appears to be an error on that site, since the PubMed record and the Science editorial summary both resolve to the abdominal RADAR paper, but I note it because it is circulating.
Sources
10 of 10 linked to recordsAlibaba DAMO Academy official GitHub organization page, listing `damo-radar` as a public repository under Apache-2.0, last updated Sep 18 2026
Hugging Face organization `radar-generalist`, hosting the RADAR model, README and `RADAR-auxiliary-data`, updated four days before retrieval
Zenodo deposit, "RADAR: An Expert-Level Generalist AI for Abdominal CT Diagnosis," source code archive linking the GitHub and Hugging Face releases
Science, "In Science Journals" editorial summary of Zhang et al. RADAR
PubMed citation record for "An expert-level generalist AI for abdominal CT diagnosis" (PMID 42752131)
ClinicalTrials.gov NCT07040358, "Rapid Abdominal Diagnosis With AI & Radiology"
EurekAlert release, "Introducing RADAR, a generalist AI tool for abdominal CT diagnosis"
South China Morning Post, "Alibaba open-sources medical AI model that can detect cancer and nearly 150 conditions"
GitHub issue #1 on `alibaba-damo-academy/damo-radar`, opened Sep 18 2026
Poniak Times and ascendants.in, on the weights licence