Alibaba open-sources RADAR to detect 146 findings in abdominal CT scans

Why it matters

RADAR gives medical-AI teams an inspectable foundation spanning 146 abdominal findings. Its breadth could shorten development cycles, but regulation and prospective clinical evidence remain the gatekeepers.

Qi Zhang and Jianpeng Zhang were among the researchers spanning Alibaba's DAMO Academy, Zhejiang University and affiliated hospitals who built an AI model that screens abdominal CT scans for 146 clinical findings, then released its code and checkpoints for other researchers to test.

The model, called RADAR, analyzes contrast-enhanced CT scans across 18 organs and anatomical structures. Alibaba published the open-source repository under the Apache 2.0 license on September 18th, one day after the team's paper appeared in Science.

The South China Morning Post's report described the model as detecting cancer and nearly 150 conditions. The underlying study evaluates 146 radiological findings, a category that includes cancers alongside other diseases and abnormalities.

The distinction matters. RADAR is a broad abdominal imaging model, rather than a detector for 146 different cancers. Its reported value lies in covering many possible findings from one scan, closer to the range a radiologist must consider during routine work.

Training on reports instead of a hand-labeled disease list

The researchers trained RADAR on 424,911 contrast-enhanced abdominal CT examinations and more than 15 million anatomy-aware image-text pairs. The system learns from clinical reports tied to scans, reducing its dependence on manually annotated datasets built separately for each disease.

That training strategy reflects the long-running focus of both lead researchers. Qi Zhang is a researcher at the First Affiliated Hospital of Zhejiang University School of Medicine whose work includes AI for cancer diagnosis. Jianpeng Zhang, a senior algorithm researcher at Alibaba DAMO Academy, previously conducted research at Zhejiang University and the Australian Institute for Machine Learning. His work centers on medical-image analysis, multimodal learning and vision-language pretraining.

Their bet is that clinical reports already contain much of the supervision needed to train broader medical-imaging models. RADAR divides CT volumes into anatomical regions and associates those regions with descriptions from radiology reports. This allows one model to learn patterns spanning common and uncommon findings without requiring clinicians to draw or label every example by hand.

In evaluations covering nearly 40,000 real-world examinations, the study reported a mean area under the curve, or AUC, of 0.913 across the 146 findings. AUC measures how reliably a model separates positive and negative cases across different thresholds. It is not the percentage of patients diagnosed correctly.

An AAAS summary of the study said the best competing vision-language model in the researchers' evaluation recorded a mean AUC of 0.776. RADAR reached 0.895 in cohorts from eight external medical centers and 0.904 across more than 27,000 emergency CT cases, according to the same summary.

Those results support the team's claim that RADAR generalizes beyond its core training set. Independent prospective studies would still need to establish its performance when inserted into routine hospital workflows, where patient populations, scanners, reporting standards and disease prevalence can shift the balance between missed findings and false alarms.

The radiologist study points toward assistance, not autonomy

The researchers also compared RADAR with 26 radiologists from multiple hospitals. The model's average accuracy exceeded that of 23 participants, according to the Science paper.

The more consequential test placed RADAR beside the clinicians. With AI assistance, the radiologists improved sensitivity by about 10%, meaning they missed fewer positive findings, while cutting reading time by more than 30%.

That result gives RADAR a plausible role as a second reader or worklist assistant. It does not establish that the model can operate independently, improve patient outcomes or safely handle every abdominal CT without a physician's review. The team's description of RADAR as the world's first "expert-level generalist medical imaging model" remains its own characterization, rather than a regulatory designation.

Alibaba has published the training and inference code, preprocessing tools and pretrained checkpoints. The repository also includes instructions for evaluating RADAR against the external MERLIN dataset. Researchers can inspect and reproduce more of the work than they could with a closed clinical model, although open weights are separate from authorization to use a system in patient care.

That gap is where medical AI becomes a business rather than a benchmark. In the United States, Aidoc has an FDA-cleared comprehensive abdominal-CT triage product, while a2z Radiology AI has clearance to triage seven urgent abdomen-pelvis findings. The FDA's AI-enabled medical device list provides the relevant regulatory context. Those products target triage and notification, while RADAR covers a much broader set of findings in the research setting.

The open release gives universities, hospitals and medical-AI developers a common foundation to evaluate, adapt and potentially build upon. It remains separate from authorization to use RADAR in patient care.

Open weights widen the test surface

DAMO Academy was established in 2017 and has since developed screening systems for pancreatic, colorectal and stomach cancers, along with aortic disease and other conditions. The group's earlier projects often concentrated on individual diseases or screening programs. RADAR packages that work into a generalist architecture intended to read across the abdomen.

The release also fits Alibaba's broader push to distribute models rather than keep its AI research behind internal products. RuntimeWire reported on September 18th that Alibaba had released Qwen3.8-Omni-Flash for text, image, audio and video work. RADAR applies a similar distribution strategy to a field where the cost of error is considerably higher and the route from repository to deployment runs through clinical validation and regulation.

For Zhang and the wider research group, open-sourcing RADAR moves the work beyond a paper comparison. Outside teams can probe performance by disease, demographic group, scanner and institution. That scrutiny will determine whether the model's breadth survives contact with hospitals beyond the study and whether generalist medical imaging can progress from a research category into dependable clinical infrastructure.