September 19, 2026
AI for CT scans goes open source: RADAR checks 146 findings
RADAR posted a mean AUC of 0.913 across 146 CT findings versus 0.776 for the best competing vision-language model on September 17, 2026. On September 18, Alibaba DAMO Academy released the code under Apache License 2.0 and documentation for demo inference on one GPU. The model was trained on 424 911 contrast-enhanced abdominal CT studies, 1.5 million image-text pairs, and more than 15 million anatomically grounded pairs.
Across more than 27 000 emergency CT scans, RADAR achieved an AUC of 0.904, although it was not trained specifically on emergency data. In external testing at 8 medical centers, it achieved an AUC of 0.895. In a study involving 26 radiologists, RADAR assistance increased diagnostic sensitivity by about 10%.
From paper to launch. On September 17, AAAS published the RADAR results in a Science paper. On September 18, the code appeared in the Alibaba DAMO Academy repository on GitHub.
For demo inference, create a conda environment with Python 3.10 and install dependencies with `pip install -r requirements.txt`. Then obtain the `ckpt/checkpoint_radar_pretrain.pth` checkpoint and `bert-base-chinese`, then the documentation instructs you to move to `RADAR_inference` and run `python inference_demo.py`. The documentation lists A100 and H20 GPUs.
On September 18, the code and demo will be available, while the published metrics cover abdominal CT, emergency studies, and tests at 8 medical centers.
Source
