MedMLX/synthseg-mlx

SynthSeg for MLX

Native float32 MLX weights for SynthSeg brain MRI and CT segmentation, cortical parcellation, and quality control. The files are converted from the original Keras checkpoints distributed by FreeSurfer; they have not been retrained. Convolution kernels use MLX's (out_channels, depth, height, width, in_channels) layout.

File	Purpose
synthseg_2.0.safetensors	Standard or fast segmentation
synthseg_robust_2.0.safetensors	Robust segmentation cascade
synthseg_1.0.safetensors	SynthSeg 1.0 segmentation
synthseg_parc_2.0.safetensors	Cortical parcellation
synthseg_qc_2.0.safetensors	Quality control

Every variant uses the parcellation and QC files, even when parcellation is disabled. On Apple silicon with MLX and Metal, install the runtime:

uv pip install "synthseg-mlx>=0.3.4" huggingface-hub

Then download the weights and run:

from huggingface_hub import snapshot_download
from synthseg_mlx.medmlx import run

weights = snapshot_download(
    "MedMLX/synthseg-mlx",
    revision="b4fa99020571cd5921f2205834dd5028fa804046",
    allow_patterns=["synthseg_2.0.safetensors", "synthseg_parc_2.0.safetensors",
                    "synthseg_qc_2.0.safetensors"],
)
result = run(image_path="brain.nii.gz", weights_path=weights, output_dir="out")

For robust=True or v1=True, download the corresponding segmentation file instead of synthseg_2.0.safetensors. See the runtime documentation for inputs and outputs.

Validation and use

The converted weights were compared against the original H5 files. On one anatomical test image, standard, fast, robust, v1, and v1-fast runs produced identical masks, output grids, volume tables, and QC tables. This is a format conversion check, not a clinical or full-population validation. Research use only; outputs require human review.

Please cite the original SynthSeg papers.

License

The weights are redistributed under the FreeSurfer Software License, reproduced in LICENSE-FREESURFER.txt and below. The converted files are modified forms of the original weights.

FreeSurfer Software License

All or portions of this licensed product (such portions are the "Software") have been obtained under license from The General Hospital Corporation and are subject to the following terms and conditions:

== PART B. DOWNLOADING AGREEMENT ==
=== License from MGH with Right to Sublicense ("Software License") ===

 1. As used in this Software License, "you" means the individual downloading and/or using, reproducing, modifying, displaying and/or distributing the Software and the institution or entity which employs or is otherwise affiliated with such individual in connection therewith.  The General Hospital Corporation, Boston MA, USA ("MGH") hereby grants you, with right to sublicense, with respect to MGH's rights in the software, and data, if any, which is the subject of this Software License (collectively, the "Software"), a royalty-free, non-exclusive license to use, reproduce, make derivative works of, display and distribute the Software, provided that: (a) you accept and adhere to all of the terms and conditions of this Software License; (b) in connection with any copy of or sublicense of all or any portion of the Software, all of the terms and conditions in this Software License shall appear in and shall apply to such copy and such sublicense, including without limitation all source and executable forms and on any user documentation, prefaced with the following words: "All or portions of this licensed product (such portions are the "Software") have been obtained under license from The General Hospital Corporation and are subject to the following terms and conditions:" (c) you preserve and maintain all applicable attributions, copyright notices and licenses included in or applicable to the Software; (d) modified versions of the Software must be clearly identified and marked as such, and must not be misrepresented as being the original Software; and (e) you consider making, but are under no obligation to make, the source code of any of your modifications to the Software freely available to others on an open source basis.

 2. The license granted in this Software License includes without limitation the right to (i) incorporate the Software into proprietary programs (subject to any restrictions applicable to such programs), (ii) add your own copyright statement to your modifications of the Software, and (iii) provide additional or different license terms and conditions in your sublicenses of modifications of the Software; provided that in each case your use, reproduction or distribution of such modifications otherwise complies with the conditions stated in this Software License.

 3. This Software License does not grant any rights with respect to third party software, except those rights that MGH has been authorized by a third party to grant to you, and accordingly you are solely responsible for (i) obtaining any permissions from third parties that you need to use, reproduce, make derivative works of, display and distribute the Software, and (ii) informing your sublicensees, including without limitation your end-users, of their obligations to secure any such required permissions.

 4. The Software has been designed for research purposes only and has not been reviewed or approved by the Food and Drug Administration or by any other agency. YOU ACKNOWLEDGE AND AGREE THAT CLINICAL APPLICATIONS ARE NEITHER RECOMMENDED NOR ADVISED. Any commercialization of the Software is at the sole risk of the party or parties engaged in such commercialization. You further agree to use, reproduce, make derivative works of, display and distribute the Software in compliance with all applicable governmental laws, regulations and orders, including without limitation those relating to export and import control.

 5. The Software is provided "AS IS" and neither MGH nor any contributor to the software (each a "Contributor") shall have any obligation to provide maintenance, support, updates, enhancements or modifications thereto. MGH AND ALL CONTRIBUTORS SPECIFICALLY DISCLAIM ALL EXPRESS AND IMPLIED WARRANTIES OF ANY KIND INCLUDING, BUT NOT LIMITED TO, ANY WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL MGH OR ANY CONTRIBUTOR BE LIABLE TO ANY PARTY FOR DIRECT, INDIRECT, SPECIAL, INCIDENTAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY ARISING IN ANY WAY RELATED TO THE SOFTWARE, EVEN IF MGH OR ANY CONTRIBUTOR HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES. TO THE MAXIMUM EXTENT NOT PROHIBITED BY LAW OR REGULATION, YOU FURTHER ASSUME ALL LIABILITY FOR YOUR USE, REPRODUCTION, MAKING OF DERIVATIVE WORKS, DISPLAY, LICENSE OR DISTRIBUTION OF THE SOFTWARE AND AGREE TO INDEMNIFY AND HOLD HARMLESS MGH AND ALL CONTRIBUTORS FROM AND AGAINST ANY AND ALL CLAIMS, SUITS, ACTIONS, DEMANDS AND JUDGMENTS ARISING THEREFROM.

 6. None of the names, logos or trademarks of MGH or any of MGH's affiliates or any of the Contributors, or any funding agency, may be used to endorse or promote products produced in whole or in part by operation of the Software or derived from or based on the Software without specific prior written permission from the applicable party.

 7. Any use, reproduction or distribution of the Software which is not in accordance with this Software License shall automatically revoke all rights granted to you under this Software License and render Paragraphs 1 and 2 of this Software License null and void.

 8. This Software License does not grant any rights in or to any intellectual property owned by MGH or any Contributor except those rights expressly granted hereunder.

== PART C. MISCELLANEOUS ==

This Agreement shall be governed by and construed in accordance with the laws of The Commonwealth of Massachusetts without regard to principles of conflicts of law. This Agreement shall supercede and replace any license terms that you may have agreed to previously with respect to FreeSurfer.

Citation

If you use MedMLX, please cite the MedMLX paper:

@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}
}