Generalist vision-language model for contrast-enhanced abdominal CT that learns organ-level findings from radiology reports, not manual labels.
No providers recorded yet. Browse all providers
A contrast-enhanced abdominal CT captures the liver, pancreas, kidneys, spleen, bowel and vessels in one volume, and the radiologist's report describes each of them in turn. Most radiology AI ignores that structure: a separate supervised detector is trained for each disease, each one needing its own hand-annotated dataset and each blind to everything outside its target. Aligning the whole scan to the whole report, as general vision-language models do with photos and captions, fails for a different reason: pathology in a CT volume is small and local, so a sentence about a lesion in one organ gets matched to hundreds of slices of normal tissue.
RADAR (released as DAMO-RADAR) addresses this by working organ by organ. Each CT volume is segmented into anatomical structures, each report is split into the sentences that describe those structures, and the model learns to match an organ's image features to its own description. The result is a single generalist model that reads one abdominal CT for 146 imaging findings across 18 anatomical structures, trained on the clinical reports hospitals already write rather than on manual labels.
RADAR was developed by Alibaba DAMO Academy with the First Affiliated Hospital, Zhejiang University School of Medicine, with corresponding authors Ling Zhang and Tingbo Liang, and published in Science in September 2026. Collaborators include Zhejiang University, Hupan Lab, and a network of regional hospitals in Zhejiang and beyond.
The vision encoder is a 3D convolutional UNet-style network that produces three levels of image features and an anatomical segmentation over 36 merged structures; the text encoder is BERT, initialized from bert-base-chinese for the original Chinese-language reports and from bert-base-uncased for English. Training is built on the LAVIS framework: image and text features are projected to a shared 256-dimensional space and trained with an image-text contrastive loss per organ, restricted to organs fully contained in the scan field. The released configuration trains for 30 epochs on 24 A100 or H20 GPUs with a total batch size of 48.
The pretraining corpus, RAD-CT, comprises more than 400,000 contrast-enhanced abdominal CT exams with paired clinical reports. In the paper, RADAR was evaluated on internal and external cohorts from multiple centers and a range of clinical scenarios. The public repository reproduces a zero-shot external test: the RAD-CT checkpoint, applied unchanged to the Stanford Merlin abdominal CT test set, reaches a mean AUC of 0.883 across 21 findings. RADAR+ trained from scratch on the Merlin training set reaches an anatomy-level AUC of 0.888, and the RAD-CT checkpoint fine-tuned on that set reaches 0.918.
RADAR is aimed at comprehensive reading of routine contrast-enhanced abdominal CT: flagging findings across the liver, biliary tree, pancreas, kidneys, spleen, bowel, vasculature and lower chest in a single pass, and acting as a second reader that directs a radiologist's attention. In the paper, RADAR's average performance exceeded that of 23 of the 26 radiologists it was compared with, and in assisted reading it increased radiologists' disease-detection sensitivity by about 10%. For researchers, the released code and checkpoints support zero-shot evaluation on new cohorts, fine-tuning on public datasets such as Merlin, and reuse of the anatomy-wise pipeline on their own image-report archives. The developers state that the model is intended for research use only and that prospective clinical studies are required before clinical deployment.
RADAR is a large-scale demonstration that organ-level vision-language pretraining on routine hospital archives can produce a single abdominal CT model covering well over a hundred findings, in contrast to the one-disease-one-model pattern of earlier radiology AI. It sits alongside other 3D CT vision-language models in the catalog such as Merlin, CT-CLIP and RadFM. The code, pretrained and Merlin-trained checkpoints, and processed Merlin masks are public under a CC BY-NC-SA 4.0 license, which permits research use but not commercial deployment. The RAD-CT training data itself is not released, and the model is limited to contrast-enhanced abdominal CT.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.