https://atlas.rsna.org/schemas/2025-11/model.json

TotalSegmentator

https://www.github.com/wasserth/TotalSegmentator

jakob.wasserthal@usb.ch

Authors declared no funding for this work.

The Ethics Committee Northwest and Central Switzerland approved the ethics waiver for this retrospective study (EKNZ BASEC Req-2022–00495).

This model provides robust and accurate segmentation of 104 anatomic structures on CT images, including 27 organs, 59 bones, 10 muscles, and 8 vessels. It was trained on a diverse real-world dataset and is publicly available as a Python package.

The model uses the nnU-Net framework, a U-Net-based implementation that automatically configures hyperparameters. Two models were trained: one with 1.5-mm isotropic resolution and another with 3-mm isotropic resolution.

The model and its training data are publicly available. The model is provided as a pretrained Python package (https://www.github.com/wasserth/TotalSegmentator) and the annotated dataset (https://doi.org/10.5281/zenodo.6802613) is freely downloadable.

Improves quality of radiologic reports, reduces radiologist workload, enables extraction of advanced biomarkers, automatic detection of abnormalities, and quantification of tumor load. Useful for organ volumetry, disease characterization, and surgical or radiation therapy planning.

Diagnosis

Fully automatic segmentation; optional manual refinement possible.

Automatic and robust segmentation of 104 major anatomic structures (27 organs, 59 bones, 10 muscles, 8 vessels) on body CT images for applications such as organ volumetry, disease characterization, and surgical or radiation therapy planning in any clinical setting.

Whole-body CT images with 1.5-mm or 3-mm isotropic resolution.

The model is available as a pretrained Python package. It requires less than 12 GB of RAM and does not require a GPU, allowing it to run on a normal laptop. Users should be aware of typical failure cases such as missing small parts of the colon or iliac arteries and mixing up neighboring vertebrae and ribs.

Male patients were over-represented in the study dataset. Typical failure cases include missing small parts of the colon or iliac arteries and mixing up neighboring vertebrae and ribs.

Use the 1.5-mm model for highest accuracy when resources allow; use the 3-mm model for limited RAM/GPU environments. Review outputs in known failure scenarios (e.g., ribs/vertebrae adjacency, small bowel/vascular segments).

The nnU-Net framework is self-configuring, which enhances reproducibility by automatically adapting hyperparameters based on dataset characteristics.

The model has low technical requirements, needing less than 12 GB of RAM and no GPU, making it runnable on a normal laptop.