Alibaba’s research arm has released something rare in medical artificial intelligence: a powerful diagnostic model that anyone can inspect, improve, and run. On September 18, 2026, the DAMO Academy open-sourced DAMO RADAR, a vision-language model that reads abdominal CT scans and flags cancer along with 145 other conditions.
The results behind the release are striking. Detailed in a peer-reviewed study published in Science, the model achieved expert-level accuracy across nearly 40,000 real-world exams and outperformed 23 of 26 radiologists in a head-to-head comparison. Here is what DAMO RADAR actually does, how accurate it is, and why open-sourcing it matters as much as the numbers.
What Is DAMO RADAR?
DAMO RADAR is a generalist medical imaging model built to analyze contrast-enhanced computed tomography (CT) scans of the abdomen. The name stands for Rapid Abdominal Diagnosis with AI and Radiology.
It was developed by Alibaba’s DAMO Academy together with Zhejiang University’s First Affiliated Hospital and other partner hospitals. Rather than targeting a single disease, the model covers 146 clinical findings across 18 abdominal organs and anatomical structures, including the liver, pancreas, gallbladder, kidneys, spleen, stomach, and intestines.
One important distinction gets lost in most headlines: those 146 items are radiological findings, not 146 different cancers. The findings span malignant tumors, chronic conditions such as fatty liver disease, and acute problems such as appendicitis. Cancers are among the conditions the model can flag, but they are a subset of its full scope.
The release continues Alibaba’s growing track record in medical AI. The company first drew global attention in this field in 2020, when researchers at DAMO Academy built a system that could detect coronavirus in seconds from CT scans with 96 percent accuracy.
How DAMO RADAR Learned to Read CT Scans
Most medical imaging AI to date has been trained the hard way. Developers pay doctors to annotate scans slice by slice, marking tumors and abnormalities by hand. The process is expensive, and each finished model typically handles just one disease or organ.
DAMO RADAR took a different path. Instead of manual annotation, the team used vision-language learning, training the model on pairs of CT scans and the clinical reports radiologists had already written about them. In total, the model trained on 424,911 contrast-enhanced abdominal CT examinations, producing more than 15 million anatomy-aware image-text pairs.
Two technical innovations made this approach work:
- Organ-level fine-grained alignment: the model breaks each 3D CT volume into anatomical units, then aligns each organ with the matching description in the radiology report, mimicking how radiologists read organ by organ.
- Adaptive contrastive modeling: the training process adjusts how it compares cases based on medical knowledge, so two patients with healthy livers are recognized as similar rather than treated as unrelated samples.
DAMO researchers noted that CT data signals are comparatively sparse, which makes standard vision-language training unreliable, and that the organ-by-organ alignment strategy is what allowed the model to learn from clinical reports without additional manual labeling.
How Accurate Is DAMO RADAR?
Medical AI performance is usually reported as AUC, or area under the curve. A score of 0.5 means the model is no better than chance at telling sick patients from healthy ones; 1.0 means perfect discrimination. DAMO RADAR’s headline number is a mean AUC of 0.913, but the fuller validation matrix is more revealing.
| Validation Setting | Scale | Mean AUC |
|---|---|---|
| Internal real-world cohort | ~39,000 exams, 146 findings | 0.913 |
| External validation, 8 medical centers | 24,000+ exams | 0.874 to 0.912 |
| Four cancers vs. pathology gold standard | Liver, pancreatic, gastric, colorectal | 0.891 to 0.984 |
| Emergency CT scans, outside training focus | ~27,000 cases | 0.904 |
The external validation line matters most. A model can look excellent on the hospital data it learned from and collapse on scans from other institutions. Across eight external medical centers, DAMO RADAR held scores between 0.874 and 0.912, a modest drop that suggests real generalization.
Two further results stand out. When tested against pathology, the gold standard that confirms cancer through tissue analysis, the model scored between 0.891 and 0.984 for liver, pancreatic, gastric, and colorectal cancers. And on roughly 27,000 emergency CT scans, a setting the model was not specifically trained for, it still reached an AUC of 0.904.
How DAMO RADAR Performed Against Human Radiologists
The Science study included a reader test with 26 radiologists from 14 hospitals. On average, the model’s independent diagnostic accuracy exceeded that of 23 of the 26 physicians.
The collaboration results may matter more. When radiologists worked with AI assistance:
- Diagnostic sensitivity rose by about 10 percent, meaning fewer missed diagnoses.
- Average reading time dropped by more than 30 percent.
- Junior radiologists performed at the level of senior physicians.
AI models besting expert readers is not without precedent; back in 2020, a breast cancer AI outperformed expert radiologists at reading mammograms. What sets DAMO RADAR apart is breadth: it competes with specialists across an entire region of the body rather than a single cancer type.
That finding about junior doctors performing at senior level hints at where the technology could have the biggest impact. The goal is not to replace radiologists but to level up diagnostic quality in hospitals that struggle to hire experienced specialists.
What Open-Sourcing Means for Healthcare
On September 18, 2026, the day after the Science paper appeared, the team published the RADAR code on GitHub under the Apache 2.0 license, with pre-trained model checkpoints released on Hugging Face and an archived copy on Zenodo. The repository includes training and inference code, preprocessing tools, and documentation for reproducing the results, and it had already drawn more than 180 stars from the developer community within a day.
Most medical AI is the opposite of this. Commercial imaging models are proprietary black boxes, their training data and inner workings hidden from the hospitals that deploy them. The open release changes what independent teams can do:
- Reproduce and audit the published accuracy claims instead of taking them on trust.
- Fine-tune the model for local populations, scanners, and disease patterns.
- Adapt the training approach to other organs and imaging types, such as chest CT or MRI.
- Give smaller hospitals and research groups access to expert-level imaging AI without licensing fees.
The researchers argue the underlying method generalizes beyond abdominal CT. Senior DAMO scientists told Chinese media that the vision-language training paradigm could extend to other forms of medical imaging, potentially accelerating the arrival of generalist clinical AI.
Open access does not mean easy deployment. The repository’s documentation states that inference can run on a single NVIDIA A100 or H20 data center GPU, with multi-GPU setups recommended for large-scale evaluation. Reproducing training from scratch is a different scale entirely: the team reports using 24 GPUs (A100 or H20) with a total batch size of 48. The released assets also include a pretrained UNet vision branch, BERT text encoders, and anatomical masks generated by TotalSegmentator.
A second nuance sits in the licensing. The code carries the permissive Apache 2.0 license, but the model weights distributed on Hugging Face are marked under a Creative Commons BY-NC-SA license, which restricts commercial use. Hospitals or companies looking to deploy the model commercially would need a separate arrangement with the developers.
Limitations: What DAMO RADAR Cannot Do
Impressive AUC scores do not make DAMO RADAR a finished clinical product. Several caveats deserve as much attention as the results:
- It is a research release, not an approved medical device. No regulatory clearance for clinical diagnosis has been announced, and the model is not licensed for unsupervised use on patients.
- Vision-language models can hallucinate. A 2026 review of vision-language models in diagnostic imaging found factual errors in roughly 22 percent of AI-generated radiology reports, with accuracy declining further on complex cases.
- Bias and generalizability remain open questions. Research on AI in radiology warns that models trained on limited demographics can perform worse on underrepresented populations, and RADAR’s training data came largely from Chinese partner hospitals.
- Human oversight is still the design. The study’s strongest results came from radiologists working with the model, not from the model operating alone.
Open-sourcing actually helps address some of these gaps, since external teams can now stress-test the model on populations and scanners its creators never used. But availability on GitHub is not the same as readiness for a hospital reading room.
Frequently Asked Questions
What does DAMO RADAR stand for?
Rapid Abdominal Diagnosis with AI and Radiology. The name reflects its focus on abdominal CT scans.
Is DAMO RADAR available to the public?
Yes. The code is on GitHub under an Apache 2.0 license, and model checkpoints are on Hugging Face. Running it requires technical setup, including a Python environment, so it currently targets researchers and developers rather than consumers.
Can DAMO RADAR replace radiologists?
No. It outperformed most radiologists in the study, but it is designed as an assistive tool. The largest measured gains came from human-AI collaboration: fewer missed diagnoses and faster reads, with doctors still making the final call.
What cancers can DAMO RADAR detect?
Cancers are among its 146 evaluated findings. Validated against pathology, the model scored AUCs of 0.891 to 0.984 for liver, pancreatic, gastric, and colorectal cancers.
Is DAMO RADAR approved for clinical use?
No clinical approval has been announced. As of September 2026, it is a research model released for study, validation, and further development.
What hardware do you need to run DAMO RADAR?
Per the repository documentation, inference can run on a single NVIDIA A100 or H20 data center GPU, though multi-GPU setups speed up large-scale work. Reproducing training from scratch is far beyond consumer hardware: the developers used 24 such GPUs. The model weights also carry a non-commercial license, separate from the Apache 2.0 code license.
Conclusion
DAMO RADAR represents a shift in medical imaging AI: from narrow, single-disease tools trained on expensive hand-annotated data toward generalist models that learn from the scans and reports hospitals already produce. Its accuracy across 146 findings, its holdout performance at external hospitals, and its assistive gains for radiologists all point to technology built for real clinical workflows.
The open-source release may prove just as important as the science. Medical imaging keeps evolving at pace, from the world’s first full-body scanner to AI models that interpret what the scans show. With DAMO RADAR’s code, checkpoints, and documentation all public, the next milestones to watch are independent, multi-country validation and the regulatory decisions that will determine whether this model reaches the bedside.
