Alibaba's research arm has open-sourced an artificial intelligence model that can identify close to 150 abdominal conditions — including cancers — by reading computed tomography (CT) scans, the latest step in the Chinese technology group's expanding medical AI program.
The model, called Damo Radar, was developed by Damo Academy, Alibaba's research institute, and described in a study published Thursday in the journal Science, according to the South China Morning Post. The institute announced the open-source release on Friday. Unlike closed commercial systems, an open-source model can be downloaded, inspected and adapted by hospitals and researchers — a distribution model that could speed adoption in clinics that could never license a proprietary diagnostic tool. For more context on this story, see our ongoing more AI stories.
A Generalist Model for Medical Imaging
Damo Radar is a vision-language model designed to analyze contrast-enhanced CT scans covering 18 abdominal organs and identify a broad range of diseases and other abnormalities, including malignant tumors, the institute said. Its creators describe it as "the world's first expert-level generalist medical imaging model" — a single system intended to read across many different findings rather than a narrow tool trained to spot one disease.
That generalist ambition sets it apart from most medical AI systems deployed so far, which have typically been built and validated for a single task, such as detecting one type of tumor. A single model that spans 146 distinct clinical findings, if it holds up outside the lab, could change the economics of deploying AI in radiology departments, where stitching together a dozen single-purpose tools has proven impractical.
The model was trained on CT scans paired with clinical reports, an approach that teaches the system to connect what appears on an image with the language clinicians use to describe it. The research team said the training method could eventually be extended to other types of medical imaging, according to the Post.
Tested Against Working Radiologists
The headline numbers come from evaluation on nearly 40,000 real-world examinations. Across 146 clinical findings, Damo Radar achieved an average area under the curve (AUC) of 0.913, a statistical measure of diagnostic performance in which 1.0 represents perfect accuracy, the study reported.
The comparison with human readers will draw the most scrutiny. In a study involving 26 radiologists from multiple hospitals, the model's average accuracy exceeded that of 23 of the participants, according to the research. And when radiologists worked with the model's assistance, they improved their ability to prevent missed diagnoses by 10 percent while cutting the time required per case by more than 30 percent, Damo Academy said.
The research involved several institutions alongside Alibaba, including a hospital affiliated with Zhejiang University. Publication in Science, one of the most selective scientific journals, gives the results a level of external vetting that vendor-sponsored benchmarks often lack — though real-world performance always depends on the scans, equipment and patient populations a hospital actually sees.
Why Open-Sourcing Matters
Releasing the weights openly is a notable choice for medically sensitive technology. Proprietary vendors typically keep medical models closed, citing regulation and safety. Open release allows independent researchers to probe failure modes, lets hospitals run the model on their own infrastructure without shipping patient data to an outside vendor, and gives smaller facilities access to capabilities otherwise reserved for large hospital systems.
It also means Alibaba exercises less control over how the model is used. Open availability is not clinical approval: regulators in each jurisdiction will still decide whether and how the model can be deployed on patients, and hospitals will need to validate it on their own equipment and populations.
One caveat worth noting: the South China Morning Post is owned by Alibaba, a fact the paper discloses in its own report. The publication of the underlying study in Science provides independent confirmation of the research itself, but the framing of Friday's announcement inevitably comes from the company's side.
Alibaba's Broader Medical AI Push
Damo Radar is not Alibaba's first move into diagnostic imaging. Damo Academy has spent recent years developing AI screening tools to help doctors identify pancreatic, stomach and colorectal cancers as well as aortic dissections, according to the Post. In April, the division released a model called Coca, co-developed with institutions including Guangdong General Hospital, which the company said was more sensitive than radiologists in spotting early-stage colorectal cancer on CT scans.
The progression from single-disease screeners toward a generalist model reflects a broader bet: that large vision-language models trained on paired images and reports can absorb pattern-recognition work that currently requires scarce specialist time. Radiologist shortages are acute in many health systems, and screening programs generate far more scans than specialists can read with equal attention.
What Comes Next
An open-source release is a beginning, not a deployment. Hospitals considering the model will need local validation on their own scans, and clinical use will depend on regulatory review country by country. But the evidence published this week — nearly 40,000 real-world exams, a head-to-head comparison against working radiologists and a reported 30 percent reduction in reading time with AI assistance — gives Damo Radar a stronger foundation than most models that debut as press releases. Whether generalist medical imaging models become standard hospital tools or remain research milestones will depend on how they perform in the messier environment of everyday clinical practice.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
