musclemap-mlx

Segment 113 muscle and bone structures in MRI or CT NIfTI images with MuscleMap v1.4, running natively through MLX and Metal on Apple Silicon. Weights download automatically on first use.

Install and segment

Requires an Apple Silicon Mac, Metal and Python 3.13. With uv:

uv venv --python 3.13
source .venv/bin/activate
uv pip install musclemap-mlx

# Real MRI example from TotalSegmentator's official test data.
curl -L https://raw.githubusercontent.com/wasserth/TotalSegmentator/d89e47f73d24f5fba732a715aeb6fdc0ae34526b/tests/reference_files/example_mr_sm.nii.gz -o mri.nii.gz
musclemap-mlx segment --input mri.nii.gz --output-dir out

The command writes out/mask.nii.gz and prints JSON containing the output path, label table, device, memory use and chunking settings. To use the corresponding CT example, replace example_mr_sm.nii.gz with example_ct_sm.nii.gz.

Python

from musclemap_mlx.medmlx import run

result = run(image_path="mri.nii.gz", output_dir="out-python")
print(result["outputs"]["mask"])

MedMLX

uv pip install 'medmlx[segment]>=0.9.1,<0.10'
medmlx describe musclemap
medmlx setup musclemap
medmlx run musclemap --arg image_path=mri.nii.gz output_dir=out-medmlx

Or use its Python interface:

import medmlx

medmlx.setup_model("musclemap")
job = medmlx.run("musclemap", arguments={
    "image_path": "mri.nii.gz",
    "output_dir": "out-medmlx-python",
})
print(job["result"])

The MCP server exposes the same model through list_models, describe_model, model_readiness, prepare_model and run_model; poll get_job for setup and segmentation results. Start it with medmlx serve --root ..

Inputs and outputs

Input is one finite 3D .nii or .nii.gz MRI or CT volume. Millimetre spatial units are supported; an unspecified unit is interpreted as millimetres, matching upstream. Other declared spatial units and 4D images are rejected.

The output is an int16 mask with background 0 and the original 113 anatomical IDs, including 1101 for left levator scapulae and 8222 for right plantaris. Its shape, affine, qform, sform and spatial-unit metadata follow the input. The largest 6-connected component of each label is retained after chunk assembly. The JSON result lists every anatomical ID and name.

Each chunk is independently normalized. For integer source images, chunking also reproduces upstream's source-header NIfTI round trip before normalization. Changing chunk size can change the mask. A chunk size at least the source depth uses one whole-volume pass. Inference uses MLX on Metal; there is no CPU or PyTorch network fallback.

Weights and offline use

The approximately 105 MB float32 bundle is hosted at MedMLX/musclemap-mlx, pinned to revision a798b3073e3b2c449987c13aa193626aaacc3092. It includes the publisher configuration, original label metadata and MIT notices.

# Pre-download to the shared cache; subsequent runs can work offline.
musclemap-mlx download

# Export a local bundle, then supply it explicitly.
musclemap-mlx download --model-dir musclemap-weights
musclemap-mlx segment --input mri.nii.gz --model-dir musclemap-weights --output-dir out-local

MedMLX and this package share the MEDMLX_ASSET_CACHE cache setting. Python callers supply a local bundle with weights_path="musclemap-weights".

Qualification and license

The native implementation was compared with the pinned MuscleMap v1.4 reference on real CT and MRI examples. Reproduction commands, acceptance criteria and measurements are in docs/technical.md. These two compact examples establish implementation parity; anatomical accuracy against manual segmentations and coverage of all 113 structures were not measured.

Code and publisher weights are MIT licensed. See publisher license and attribution. The original checkpoint and configuration are published in Zenodo record 21929873.