Alibaba’s research arm has open-sourced RADAR, a vision-language model that reads contrast-enhanced abdominal CT scans and flags close to 150 distinct conditions, from liver and pancreatic cancers to fatty liver disease and acute appendicitis. The release came alongside a peer-reviewed paper in Science, which is a rarer credential than most medical AI press releases can claim.
The numbers from the paper
RADAR, short for Rapid Abdominal Diagnosis with AI and Radiology, was built by Alibaba DAMO Academy together with the First Affiliated Hospital of Zhejiang University School of Medicine and collaborators at roughly two dozen hospitals. The training set, RAD-CT, contains 424,911 contrast-enhanced abdominal CT exams and about 15 million anatomy-aware image-text pairs. Notably, the model learned from clinical reports directly rather than from manually drawn annotations, which is what made a dataset of that size practical to assemble.
Performance held up across the evaluations the paper reports:
- Mean AUC of 0.913 across 146 findings in an internal real-world cohort of 39,160 exams.
- External validation across eight centers at AUCs between 0.874 and 0.912, and 0.883 on a cross-population cohort with no fine-tuning.
- On pathology-confirmed cancers of the liver, pancreas, stomach, and colorectum, AUCs between 0.891 and 0.984.
- In a reader study with 26 radiologists from 14 centers, the model outperformed most participants, and radiologists working with it reportedly gained around 10 percent in diagnostic sensitivity while cutting reading time by about 30 percent.
Some caveats apply. The sensitivity figure does not specify absolute versus relative improvement, and the head-to-head comparison was framed by the authors themselves. It was a reader study, not a clinical trial, and no patient outcomes were measured. Third-party reproduction is still pending.
The licensing split matters
The code on GitHub is Apache 2.0. The model weights on Hugging Face carry a CC BY-NC-SA 4.0 badge, which is a non-commercial, share-alike license. Read those two facts together: a hospital IT team can study the pipeline freely, but anyone planning a commercial product on top of the released weights needs a different arrangement. This split is common in open-source medical AI and is worth checking before architecture decisions get made on the assumption that “it is open source” means unrestricted use.
Why the training approach is the deeper story
Manual annotation is the bottleneck that has slowed radiology AI for a decade. Radiologist time is expensive, labeled CT datasets are small, and most models in this space were trained to detect one disease at a time. RADAR’s approach of learning directly from the free-text reports radiologists already produce removes that constraint, and it is the main reason one model can cover 146 findings instead of 146 separate models.
That shift has operational consequences. Hospitals evaluating this stack move from managing a portfolio of single-disease tools to deploying one generalist with a consistent interface, attention maps for interpretability, and a single update path. Prior open-source radiology generalists like RadFM and Stanford’s CheXagent either cover different modalities or focus on chest X-rays, so RADAR is entering relatively open territory for abdominal CT.
Inside the model
Architecturally, RADAR pairs a 3D vision branch that ingests CT volumes with BERT-family text encoders in both Chinese and English, connected through a contrastive learning objective. The paper’s methodological contributions are organ-level fine-grained alignment, meaning the model decomposes a CT volume into anatomical units and matches each one to the relevant slice of the report text, plus an adaptive contrastive component for rarer findings. Attention maps give clinicians a view into which regions drove a prediction, which is a concrete safety feature that pure text-based medical chatbots lack.
DAMO Academy has been building toward this for a while. The group previously published several papers in Nature Medicine on AI detection of pancreatic, gastric, and colorectal cancer, each narrowly scoped. RADAR is the consolidation of that line of work into a single generalist model, and the Science paper is the visible result.
What to do with it
The checkpoint and inference code are on GitHub and Hugging Face, so anyone with a GPU and a pile of abdominal CTs can run the demo and see how the model behaves on their own data. The repo includes fine-tuning scripts, which is the sensible first step for a research group: evaluate on local data before trusting any of the paper’s numbers as transferable. And given the non-commercial weights license, check the terms before building anything beyond a research prototype.
For the broader field, the release is a benchmark of sorts. An expert-level generalist for a whole anatomical region, published in a top journal and downloadable the same day, sets a bar that other labs will be measured against. The next question is whether independent evaluations confirm the numbers, and whether hospital deployments follow the early clinical rollout the Zhejiang team is pushing for. Either way, the era of one-model-per-disease radiology AI is ending faster than most hospital procurement cycles can keep up with.