Automated whole-body muscle segmentation

MuscleMap segmentation is performed using a deep learning–based, contrast-agnostic whole-body model trained for robust muscle and bone segmentation across MRI and CT modalities.

This page explains:

  • how to run segmentation module: mm_segment
  • all available options and flags
  • whole-body vs. regional models
  • recommended workflows
  • troubleshooting tips

Tip: For the most up-to-date list of options in your installation, run:

mm_segment --help

1. Basic usage

1.1 Activate the MuscleMap environment 

conda activate MuscleMap

1.2 Run mm_segment on a single NIfTI image

First, navigate to the directory containing your input image (or provide the full path to the file):

cd /path/to/your/data

Then run:

mm_segment -i image.nii.gz

This command:

  • loads the input image
  • applies the default whole-body segmentation model
  • writes a segmentation labelmap (NIfTI) next to the input file
  • prints output paths and logs to the terminal
Warning
Running mm_segment from the command line works as expected when MuscleMap is installed in editable mode (pip install -e .).
If the package was installed using pip install ., it may be preferable to run the script directly using python mm_segment.py.

2. Input requirements

2.1 Supported image types

  • Modality: Axial MRI of any contrast (T1w, T2w, Dixon water/fat/in-phase) or CT
  • Format: NIfTI (.nii, .nii.gz)
  • Orientation: Axial orientation recommended

MuscleMap recommends storing data in BIDS format with JSON sidecars from dcm2niix.

2.2 Processing single images vs. datasets

mm_segment processes one image at a time.

For multiple images:

  • loop through images in shell/Python
  • keep consistent naming using BIDS

3. Output

mm_segment generates deep learning–based whole-body muscle segmentation outputs, including:

  • a segmentation labelmap (NIfTI) with integer labels
  • an output file saved next to the input
  • terminal logs with paths and runtime info

Visualise results in:

4. Models and regions

4.1 Whole-body model (default) 

mm_segment -i image.nii.gz

Segments:

  • trunk muscles
  • pelvic muscles
  • thigh muscles
  • leg muscles
  • neck muscles
  • selected bones

4.2 Legacy regional models

Available:

  • abdomen
  • pelvis
  • thigh
  • leg

Example:

mm_segment -i image.nii.gz -r abdomen
Warning
The legacy regional models are maintained for backward compatibility only. Active development and state-of-the-art performance are provided exclusively by the whole-body model, which achieves robust performance across all anatomical regions.

5. Command-line options

5.1 -i — input image (required) 

mm_segment -i /path/to/image.nii.gz

5.2 -s — sliding-window overlap

Controls tile overlap.

Example:

mm_segment -i image.nii.gz -s 50
Note
Lower sliding-window overlap increases inference speed but may reduce segmentation performance, particularly near tile boundaries. Higher overlap improves robustness at the cost of longer runtimes.

5.3 -r — region model 

mm_segment -i image.nii.gz -r thigh

5.4 -g — GPU/CPU selection

Controls whether inference is performed on the GPU or CPU.

Example:

mm_segment -i image.nii.gz -g Y
Note
By default, MuscleMap runs inference on the **GPU** when available. Use `-g N` to explicitly force CPU-based inference.

5.5 -h — help 

mm_segment -h

6. Example workflows

6.1 Whole-body segmentation 

mm_segment -i sub-01_water.nii.gz

6.2 Faster inference 

mm_segment -i sub-02_T2w.nii.gz -s 50

6.3 Using a regional model 

mm_segment -i sub-03_T2w.nii.gz -r abdomen

6.4 Full pipeline with metrics 

mm_segment -i image.nii.gz
mm_extract_metrics -m gmm -r wholebody -i image.nii.gz -s image_dseg.nii.gz -c 3

See the Muscle quantification and metric extraction page for details on muscle volume and fat infiltration analysis.

7. Best practices & troubleshooting

  • Always visually inspect segmentations
  • Poor image quality reduces accuracy
  • Record MuscleMap version and options in publications

If something looks wrong: