Blame for README.md

---
license: apache-2.0
library_name: mlx
pipeline_tag: image-segmentation
tags:
- mlx
- apple-silicon
- medical-imaging
- segmentation
- nnunet
- safetensors
- research
---

# TotalSegmentator v2 CT for MLX

Native MLX safetensors for TotalSegmentator v2 CT inference on Apple Silicon. The five 1.5 mm standard models are in `291`–`295`, the 3 mm fast model in `297`, the 6 mm coarse model in `298`, and the body crop model in `300`. Each folder includes nnU-Net plans and model metadata.

The weights were converted from the [official TotalSegmentator v2.0.0 weight archives](https://github.com/wasserth/TotalSegmentator/releases/tag/v2.0.0-weights) for use with [totalsegmentator-mlx](https://github.com/MedMLX/totalsegmentator-mlx). Conversion changes tensor names and layouts for MLX; the models were not retrained.

## Quick start

```bash
uv venv --python 3.13
source .venv/bin/activate

uv pip install \
  "nnunet-mlx>=0.6.4" \
  "totalsegmentator-mlx>=0.1.6" \
  huggingface-hub
hf download MedMLX/totalsegmentator-mlx --revision 6c03d6f90ccf53f224c3985cbdecf1eaa96555af --local-dir models
totalsegmentator-mlx predict -i ct.nii.gz -o mask.nii.gz --model-dir models
```

The `total` and `body` tasks are distributed by the authors under Apache-2.0. See the [upstream repository](https://github.com/wasserth/TotalSegmentator) and the included license for terms and attribution. Please cite TotalSegmentator and nnU-Net as described by the upstream authors. For research and education; outputs require human review.

## Citation

If you use MedMLX, please cite the [MedMLX paper](https://doi.org/10.64898/2026.10.28.26366762):

```bibtex
@article{sandoval2026medmlx,
  title   = {MedMLX: On-Device Medical Imaging AI for Apple Silicon},
  author  = {Sandoval, A. and Sandoval, J. and Saleem, A.},
  journal = {medRxiv},
  year    = {2026},
  doi     = {10.64898/2026.10.28.26366762}
}
```