library_name: pytorch
license: cc-by-nc-4.0
pipeline_tag: image-segmentation
tags:
- medical-image-segmentation
- ct
- lower-extremity
- body-composition
- computer-vision
LegSegNet
LegSegNet is a deep learning system for lower extremity CT tissue segmentation and body composition quantification, introduced in the paper LegSegNet: A Public Deep Learning System for Lower Extremity CT Tissue Segmentation and Quantification.
GitHub: https://github.com/mazurowski-lab/LegSegNet
Given an input CT scan, LegSegNet segments four tissue compartments: bone, skeletal muscle, subcutaneous adipose tissue (SAT), and inter- and intramuscular adipose tissue (IMAT).
The system can further convert predicted masks into quantitative measurements, including tissue area, tissue volume, CT attenuation, and tissue-volume ratios, supporting downstream medical image analysis.
Model Details
- Task: Lower extremity CT tissue segmentation
- Input: Lower extremity CT images (PNG) or volumes (NIfTI)
- Output: Multi-class segmentation mask and body composition measurements
- Labels:
- 0: Background
- 1: SAT (Subcutaneous adipose tissue)
- 2: Muscle (Skeletal muscle)
- 3: Inter/Intra (Inter- and intramuscular adipose tissue)
- 4: Bone
Usage
LegSegNet uses a pretrained nnU-Net model. For a practical end-to-end workflow including a Gradio interface, please refer to the official GitHub repository.
Installation and Inference
To run the application locally:
pip install -r requirements.txt
python app.py
File Structure
Ensure the model files are organized as follows to use the inference scripts:
model/
|-- plans.json
|-- dataset.json
|-- fold_0/
|-- checkpoint_best.pth
Citation
If you find LegSegNet useful, please cite the following manuscript:
@article{legsegnet2026,
title={LegSegNet: A Public Deep Learning System for Lower Extremity CT Tissue Segmentation and Quantification},
author={Gao, Chen and others},
journal={arXiv preprint arXiv:2605.30829},
year={2026}
}
License
This project is released under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).