TotalSegmentator is a segmentation tool that identifies over 100 anatomical structures in CT and MR images.
The tool addresses the challenge of automatically delineating major anatomical structures across diverse medical imaging data. It uses a deep learning approach trained on a broad range of CT and MR images from different scanners, institutions, and protocols, enabling it to generalize well to new images. The model is built on nnUNet and can segment structures across the full body or focus on specific anatomical regions through specialized subtasks. Beyond segmentation, the tool can predict patient characteristics including height, weight, age, and sex from imaging data.
Developers should adopt this tool for research, clinical decision support, or automated analysis pipelines that require anatomical structure identification in CT or MR images. It suits projects needing robust segmentation across heterogeneous imaging sources without requiring extensive retraining. The tool runs on CPU and GPU across Ubuntu, Mac, and Windows, with options for faster inference on resource-constrained environments. While the README notes this is not a medical device and not intended for clinical use, it is incorporated as a component in several FDA-approved products. Users can test the tool online or integrate it as a 3D Slicer extension before committing to local deployment.
Development activity shows consistent expansion of the tool's capabilities through new task variants, particularly improvements to vertebrae segmentation to reduce anatomical confusion. The project has created multiple specialized web applications for clinical measurements including organ volume assessment, aorta diameter analysis, spine reporting, and contrast phase detection, indicating active refinement toward practical medical applications. The team maintains comprehensive documentation of supported anatomical structures and provides access to substantial training datasets for transparency and reproducibility.
