wasserth/TotalSegmentator

Tool for robust segmentation of >100 important anatomical structures in CT and MR images

What it solves

Radiologists and researchers spend a lot of time manually tracing organs, bones, vessels, and other structures in CT and MRI scans. TotalSegmentator automates this step: given a scan, it outputs labeled segmentations (3D masks) for a wide range of anatomical classes, saving hours of work and making measurements like organ volume or vessel diameter quick and reproducible.

How it works

TotalSegmentator uses deep neural networks (specifically, it builds on the nnU-Net framework) that were trained on a large, diverse set of CT and MR images from different scanners and hospitals. You run it as a command-line tool: provide an input scan (NIfTI file, or a folder/ZIP of DICOM slices) and an output folder, and it writes one NIfTI file per segmented structure (or a single label file if you prefer). It supports multiple task-specific models: the default 'total' task covers 117 classes for CT, while 'total_mr' covers 50 classes for MR. There are also many subtasks for specific anatomy (e.g., lung vessels, vertebrae, liver segments, teeth, brain structures) and optional features like fast mode (lower resolution, quicker), ROI subset selection (only segment chosen classes), statistics generation (volume, mean intensity, radiomics features), and even prediction of height, weight, age, and sex from the scan.

Who it's for

This tool is aimed at anyone working with medical CT or MR images: radiologists, clinical researchers, medical physics staff, or software developers building medical imaging pipelines. It works on Ubuntu, Mac, and Windows, and can run on CPU or GPU. It is free for academic and non-commercial use under an Apache-2.0 license for many tasks (some advanced tasks require a separate license, with free academic licenses available). Note: the authors state it is not a medical device and not intended for clinical use by itself, though it has been certified as a component in several FDA-approved products.

Highlights

  • Segments 117 anatomical classes in CT and 50 in MR with the default models
  • Dozens of specialised subtasks: lung vessels, vertebrae, liver segments, teeth, brain structures, coronary arteries, and more
  • Outputs NIfTI masks per structure, or a single label map, DICOM-SEG, or DICOM RTSTRUCT
  • Generates statistics (volume, intensity, radiomics) and 3D preview images
  • Fast mode and ROI subset options to trade accuracy for speed
  • Includes a tool to search available structures and tasks online or via totalseg_info

Related

  • Project
    JoHof/lungmask

    Lungmask provides pre-trained U-net models for automated lung segmentation in CT scans, including variants that separate lobes, handle severe pathologies, and work with COVID-19 images. It is easy to use via command line or Python and is designed to work reliably even when dense tumors, effusions, or fibrosis are present.

  • Project
    ANTsX/ANTs

    A C++ library for high-dimensional medical image registration and segmentation, used to analyze brain structure and function across species and organ systems.

  • Project
    InsightSoftwareConsortium/ITK

    An open-source, cross-platform toolkit for N-dimensional scientific image processing, segmentation, and registration, primarily used for medical image analysis.

  • Project
    Joey-S-Liu/MedSAM3

    A text-guided medical image segmentation model that uses medical concepts to identify and segment targets across diverse modalities without needing bounding boxes or points.

  • Project
    mahmoodlab/TRIDENT

    A toolkit for large-scale whole-slide image processing in AI pathology, providing an end-to-end pipeline for tissue segmentation and feature extraction using numerous foundation models.