Alibaba DAMO Academy open-sources RADAR, a medical AI that reads abdominal CT scans

Alibaba Group's research division, DAMO Academy, has open-sourced a medical artificial intelligence model that reads contrast-enhanced abdominal CT scans and flags nearly 150 conditions, including several cancers. The model, called DAMO RADAR, was introduced alongside peer-reviewed research published in the journal Science on September 18, 2026, and its code was released at the same time so hospitals and academic groups can test it without paying a licence fee. DAMO Academy, based in Hangzhou, is the research arm of the Chinese e-commerce and cloud giant Alibaba Group Holding.

The system is a vision-language model, meaning it learns from images and from the written text that accompanies them. In this case the text is the radiology report a human clinician writes after reviewing a scan. By tying specific anatomical structures inside a three-dimensional CT volume to the sentences that describe them, the researchers built a model that can answer questions about many organs at once rather than hunting for a single disease. The team describes RADAR as a generalist system for abdominal imaging, and says the training method could eventually be extended to other types of medical imaging.

Abdominal CT is widely regarded as one of the hardest areas in medical imaging because a single scan can contain the liver, pancreas, stomach, kidneys, bowel, blood vessels and lymph nodes, and abnormalities in one organ can be subtle or can mimic findings elsewhere. Most clinical AI tools built so far are narrow specialists trained to detect one condition, such as a particular type of liver lesion. DAMO RADAR's designers argue that a single generalist model can cover the same ground as many of those narrow tools and do it in one pass over one study.

The headline number is a mean area under the curve of 0.913 across 146 clinical findings, measured in nearly 40,000 real-world examinations. The area under the curve, or AUC, is a standard statistical measure of diagnostic discrimination, where 1.0 represents perfect accuracy and 0.5 is equivalent to guessing. The training pipeline was far larger than the evaluation set: the researchers trained the model on 424,911 examinations containing 1.5 million image-text pairs and more than 15 million anatomy-specific pairs.

Key Facts

South China Morning Post reported on September 18 that DAMO Academy had open-sourced DAMO RADAR, describing a vision-language model that analyses contrast-enhanced CT scans covering 18 abdominal organs and identifies malignant tumours and other abnormalities. The outlet reported that the model achieved a mean AUC of 0.913 across 146 clinical findings in nearly 40,000 real-world examinations, and noted that the academy described it as the world's first expert-level generalist medical imaging model.

Yicai Global reported on September 18 that the model covers 146 clinical findings across 18 anatomical structures and could replace many single-disease AI detectors with a single system. The same report said the team compared RADAR with 26 radiologists and found its average performance exceeded that of 23 of them, and that the study showed the model could help radiologists raise disease-detection sensitivity by 10 percent while cutting the time required by 30 percent. It quoted Zhang Ling, a senior algorithm expert at DAMO Academy, saying the model has strong generalization capabilities and broad applicability.

EurekAlert, the news service of the American Association for the Advancement of Science, reported on September 17 that RADAR was trained on a dataset of 424,911 examinations containing 1.5 million image-text pairs and more than 15 million anatomy-specific pairs. The release said the model reached a mean AUC of 0.913 across 146 abdominal CT findings, against 0.776 for the best competing vision-language model, and that it scored 0.904 across more than 27,000 emergency CT cases even though it was not specifically trained on emergency data. In testing in cohorts at eight external centres, it maintained an AUC of 0.895.

Tech Times reported on September 19 that the system appeared in Science on September 18, 2026, and that its code was posted simultaneously on GitHub and HuggingFace under an Apache 2.0 licence, with model weights released separately under a CC BY-NC-SA 4.0 licence that permits research use but prohibits commercial deployment without a separate agreement. That report also described a formal reader study against 26 specialist radiologists spanning junior residents and attending-level physicians, and said AUC ranged from 0.891 to 0.984 for liver, pancreatic, gastric and colorectal cancers where pathological biopsy served as ground truth. In one cross-population cohort, RADAR reached an AUC of 0.883 without any fine-tuning.

The lead author listed on the Science paper is Qi Zhang. DAMO Academy was founded by Alibaba in 2017 and also works on the Tongyi family of AI models, XuanTie processors and the Hanguang 800 AI chip. The research group contends that abdominal CT scans are among the most complex medical images while existing AI models typically target individual diseases, which is the gap the generalist design is meant to close.

Analysis

What this really means is that the competitive frontier in medical imaging AI has moved from single-disease detectors to generalist systems, and the proof point is no longer a laboratory benchmark but a head-to-head comparison with working physicians. A mean AUC of 0.913 is not a perfect score, and a model that beats 23 of 26 radiologists on average has still lost to three of them. But the framing matters: the comparison spans junior residents and attending-level physicians, and the reported gains when RADAR is used as a second reader are operational rather than purely statistical.

The 0.776 figure for the best competing vision-language model is the more telling number for the industry. It suggests that the gap between a general-purpose medical vision-language system and a well-tuned competitor is not marginal, at least on this benchmark. If that gap holds up under independent replication, it puts pressure on the many vendors selling narrow, disease-specific detection tools into radiology departments.

The bigger picture here is that open-sourcing has become a distribution strategy in medical AI, not an act of charity. Publishing code under Apache 2.0 while keeping weights under a non-commercial licence gives Alibaba a large pool of academic testers and citations while reserving commercial deployment for negotiated agreements. That is a deliberate split, and it mirrors how other large model providers have tried to seed ecosystems without giving away the commercial upside.

There are reasons for caution. The strongest emergency-setting figure, an AUC of 0.904 across more than 27,000 cases, comes from a model that was not trained on emergency data, which is impressive but also means the evaluation is a generalization test rather than a deployment simulation. The external validation at eight centres produced a lower AUC of 0.895, which is consistent with the usual drop when a model meets new scanners, new protocols and new patient populations.

Why It Matters

Radiology is one of the few medical specialties where workload is directly tied to image volume, and abdominal CT sits at the centre of that pressure. A tool that reads 18 organs and 146 findings in one pass, and that cuts reading time by more than 30 percent while raising sensitivity by about 10 percentage points, addresses the two constraints that most limit imaging departments: how long a read takes and how many subtle findings a tired reader misses.

The reported effect on junior radiologists is arguably the most consequential claim. If residents working with RADAR assistance reach diagnostic accuracy comparable to more experienced colleagues working unaided, then the model becomes a training and staffing tool as much as a diagnostic one. That has implications for hospitals in regions with shortages of subspecialist radiologists, and for how quickly a generalist model could be absorbed into routine reading lists.

It also matters because the evaluation covers cancers where biopsy served as ground truth. For liver, pancreatic, gastric and colorectal cancers, AUC ranged from 0.891 to 0.984. Pancreatic cancer in particular is usually found late, and any reliable improvement in early detection on routine abdominal scans would change outcomes at population scale.

Next Up

The immediate question is replication. Open weights and Apache 2.0 code mean independent groups can now attempt to reproduce the 0.913 mean AUC on their own scanners and patient populations, and the licence split means most of that work will happen in academic and non-commercial settings first. The researchers have also said the training method could eventually be extended to other types of medical imaging beyond abdominal CT.

The next milestone will be whether a generalist model of this kind reaches prospective, clinical-trial-grade evidence, and whether hospitals can deploy it inside regulated workflows. Until then, DAMO RADAR stands as a strong research result with an unusually permissive release, and a clear signal that Alibaba intends to keep building in medical AI.