wasserth/TotalSegmentator

What it solves

TotalSegmentator is designed to automate the segmentation of anatomical structures in CT and MR images. It eliminates the manual effort of identifying and outlining organs, bones, and other tissues in medical imaging, providing a robust tool that works across different scanners, institutions, and protocols.

How it works

Based heavily on the nnU-Net framework, the tool uses deep learning models trained on extensive datasets of CT and MR images (including over 1,200 CT subjects and 600 MR subjects). Users can input Nifti files or DICOM slices, and the tool predicts the location and volume of specific anatomical structures based on the selected task (e.g., the default total task for 117 classes).

Who it’s for

It is primarily intended for medical researchers and radiologists who need to extract anatomical data from medical imaging for analysis, though it is explicitly noted as not being a medical device for clinical usage.

Highlights

• Broad Anatomical Coverage: Supports a vast array of structures across multiple specialized tasks, including lung vessels, vertebrae, liver segments, and brain structures.
• Cross-Modality Support: Works on both CT and MR images.
• Body Statistics: Capable of predicting height, weight, age, and sex from images.
• Flexible Deployment: Runs on Ubuntu, Mac (with MPS support), and Windows, utilizing either CPU or GPU.
• Integration: Available as a 3D Slicer extension and via various web applications for specific reports (e.g., aorta diameter or spine reports).