Medical imaging · Software
RADAR CT
Generalist vision-language model that segments 36 anatomical structures and scores 145 findings on abdominal CT.
What RADAR CT is
RADAR is a generalist vision-language artificial intelligence model designed for interpreting contrast-enhanced abdominal computed tomography (CT) examinations. It is capable of segmenting 36 anatomical structures and generating diagnostic scores across 145 clinical findings.
Trained on over 400,000 abdominal CT studies paired with 15 million anatomy-aware image–text pairs, the model learns directly from narrative clinical radiology reports without manual labelling. It provides an automated approach for both routine abdominal segmentation and complex diagnostic evaluation.
RADAR is developed for research use only and is not a medical device. Prospective clinical investigations are required before it can be applied in clinical practice.
How it works
RADAR links three-dimensional CT imaging features to radiological text by training on 15 million anatomy-aware image–text pairs drawn from more than 400,000 contrast-enhanced abdominal CT examinations. This multi-modal supervision allows the network to learn anatomical and pathological representations directly from routine clinical reports without hand-crafted annotations.
The codebase and model architectures build upon several open-source libraries, including LAVIS for vision-language modelling, nnU-Net and MONAI for medical image processing, and 3D-ResNets-PyTorch. The pipeline contains modules for image, segmentation mask, and text report preprocessing, and supports pre-training from scratch as well as fine-tuning models (such as RADAR and RADAR+) on downstream benchmarks like MERLIN.
Key facts
- Model type
- Vision-language foundation model
- Input modality
- Contrast-enhanced abdominal CT scans (NIfTI format)
- Outputs
- 36 anatomical segmentations and 145 finding scores (CSV format)
- Training volume
- 400,000+ CT examinations with 15 million image–text pairs
- Underlying frameworks
- LAVIS, nnU-Net, MONAI, 3D-ResNets-PyTorch, PyTorch
What you can use it for
- Abdominal organ segmentation
- Extracting automated voxel-level segmentations for 36 abdominal anatomical structures from contrast-enhanced CT scans.
- Clinical finding identification
- Scoring 145 distinct pathological and diagnostic findings on abdominal scans to assist imaging research.
- Model fine-tuning on custom cohorts
- Adapting pre-trained RADAR weights to specific clinical datasets or research cohorts using provided training pipelines.
- Benchmark evaluation
- Evaluating zero-shot or fine-tuned model performance on standard external datasets such as the MERLIN test set and RAD-CT.
Run it here
Ask the research agent and it runs RADAR CT as a job on a dedicated machine, then reads the results back to you. Nothing to install, and it keeps running if you close the page.
Try asking
“Read this abdominal CT with RADAR and summarise the validated findings.”
Strengths and limits
Good at
- Trained on a large corpus of over 400,000 clinical examinations and 15 million anatomy-aware image–text pairs.
- Learns directly from unstructured clinical radiology reports without requiring human manual annotation.
- Provides broad diagnostic coverage by addressing 36 anatomical segmentations and 145 clinical findings simultaneously.
- Integrates established medical imaging and vision-language tooling including MONAI, nnU-Net, and LAVIS.
Watch out for
- Restricted to research use only and cannot be deployed in clinical practice without prospective clinical trials.
- Configured specifically for contrast-enhanced abdominal CT, with no documented support for non-contrast CT or other body regions.
- Relies on specific pinned software environments, including older framework versions such as Transformers 4.25.
Questions
Can I run RADAR CT online without installing it?
Ask the research agent and it runs RADAR CT as a job on a dedicated machine, then reads the results back to you. Nothing to install, and it keeps running if you close the page.
What licence is RADAR CT under?
CC BY-NC-SA 4.0 (code and weights); research use only per the authors. The terms are published at https://github.com/alibaba-damo-academy/damo-radar.
What formats does RADAR require for input data?
The model processes volumetric CT scans in NIfTI format alongside corresponding auxiliary masks and text reports.
Can RADAR be used as an approved diagnostic tool in a hospital?
No. RADAR is not a certified medical device and is intended solely for research purposes until prospective clinical evaluations take place.
Does training require manually segmented labels?
No. RADAR was trained directly on routine clinical radiology reports paired with CT scans, eliminating the need for manual segmentation or label annotation during initial training.
What pretrained variants exist in the repository?
The project supplies pretrained weights under checkpoint_radar_pretrain.pth and supports training both RADAR and RADAR+ configurations.
How to cite it
Qi Zhang, Jianpeng Zhang, Weiwei Cao, Zilin Lu, Wanxing Chang, Haonan Ding, Cao Chen, Zhi Li, Xing Xue, Sinuo Wang, Shaoteng Zhang, Yutong Xie, Yong Xia, Qi Wu, Zhongyi Shui, Xi Li, Zhilin Zheng, Yanjie Zhou, Tony C.W. Mok, Yingda Xia, Hongkan Wang, Xianghua Ye, Tao Ma, Jie Peng, Xiaoguang Wang, Jian Ding, Yuming Gao, Huazhen Ye, Yiping Liu, Dongjie Chen, Zhaomin Ni, Jianwen Ning, Wei Zhang, Jian Liu, Chaohui Yu, Shenghong Ju, Jianfeng Zhang, Wenbo Xiao, Ling Zhang, Tingbo Liang. An expert-level generalist AI for abdominal CT diagnosis. Science 393 (6817), eaec6129 (2026). DOI: 10.1126/science.aec6129.
Written from RADAR CT's own README and project page. Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.