Researchers from the Menze Lab at DQBM report in Nature Biomedical Engineering the development of CT-RATE, a large-scale open dataset linking 25,692 3D chest CT scans with corresponding radiology reports . The dataset enables multimodal learning directly from routine clinical data.
Leveraging CT-RATE, the team created CT-CLIP, a vision-language foundation model that outperforms fully supervised approaches in multi-abnormality detection across internal and external cohorts. They further developed CT-CHAT, a multimodal AI assistant for 3D CT that generates clinically grounded answers and reports.
By releasing CT-RATE, CT-CLIP and CT-CHAT as open resources, the study establishes a foundation for scalable, generalist AI systems in 3D medical imaging.
https://doi.org/10.1038/s41551-025-01599-y