medsam2-mlx

Segment lesions or organs from a box, point or mask prompt, then track them through a volume with MedSAM2.

Docs · MedMLX

One box prompt → nodule masks propagated across neighboring CT slices.

Apple Silicon · Metal · Python 3.13 / 3.14 · uv

Quick start

uv venv --python 3.13 --seed
source .venv/bin/activate
python -m pip install "medsam2-mlx[conversion]==0.2.1" huggingface-hub

The tagged GitHub release is an alternative installation route:

repo="https://github.com/MedMLX/medsam2-mlx.git@v0.2.1"
python -m pip install "medsam2-mlx[conversion] @ git+$repo" huggingface-hub

The package contains code only. Obtain the original checkpoint from wanglab/MedSAM2 under its publisher terms and convert it locally. The MedMLX converted CTLesion bundle remains private because its weight-card terms conflict; no converted weights are distributed here. Download and convert once, then run:

from pathlib import Path

from huggingface_hub import hf_hub_download
from medmlx_core import import_mlx
from medsam2_mlx.conversion import convert_weights

mx = import_mlx()
weights = convert_weights(Path(hf_hub_download("wanglab/MedSAM2", "MedSAM2_CTLesion.pt")))
mx.save_safetensors(
    "MedSAM2_CTLesion.safetensors", {name: mx.array(value) for name, value in weights.items()}
)

from medsam2_mlx.medmlx import run

result = run(
    image_path="ct.nii.gz",
    weights_path="MedSAM2_CTLesion.safetensors",
    output_dir="out",
    box_xyxy=[120, 140, 180, 200],
    frame_idx=40,
    propagation="both",
)
print(result["outputs"])

For NIfTI, frame_idx uses the third voxel axis. Box and point coordinates use the first two stored voxel axes.

License

Apache-2.0. See third-party notices. Model weights retain their publisher terms.

For research and education. Outputs require human review.