MeshFL is an advanced framework for distributed learning in neuroimaging. Built on the MeshNet models and NVFlare, it enables federated training for 3D MRI brain segmentation across decentralized data sites, maintaining privacy and efficiency.
For more information about MeshFL, please refer to this detailed Wiki
- Federated training of the MeshNet model for 3D MRI brain segmentation.
- Supports decentralized learning across multiple sites using NVFlare.
- Automated data handling and splitting.
- Optimized GPU usage.
- Customizable training workflows with integrated Dice score evaluation.
MeshFL training and MRI segmentation outputs
To start MeshFL, please refer to this steps here
A complete demonstration of MeshFL training and inference is available:
This demo uses 150 training rounds for faster execution. For improved performance, increasing training rounds (e.g., 300–350) can yield Dice scores above 0.93.
The demo includes:
- Training convergence visualization for 150 training rounds.
- Learning rate scheduling
- Final segmentation results
- MeshFL v2.0.0 has been released
- MeshFL paper was published in the Journal of Open Source Software (JOSS) on August 26, 2026.
The MeshFL paper, corresponding to MeshFL v2.0.0, was published on August 26, 2026 in the Journal of Open Source Software (JOSS):
For APA style, please cite the paper as:
Masoud, M., Reddy, P., Panta, S., & Plis, S. (2026). MeshFL: A Federated MeshNet Framework for 3D Brain MRI Segmentation. Journal of Open Source Software, 11(124), 9357. https://doi.org/10.21105/joss.09357
For BibTeX:
@article{Masoud2026,
doi = {10.21105/joss.09357},
url = {https://doi.org/10.21105/joss.09357},
year = {2026},
publisher = {The Open Journal},
volume = {11},
number = {124},
pages = {9357},
author = {Masoud, Mohamed and Reddy, Pratyush and Panta, Sandeep and Plis, Sergey},
title = {MeshFL: A Federated MeshNet Framework for 3D Brain MRI Segmentation},
journal = {Journal of Open Source Software}
}We welcome contributions to MeshFL! Whether it's bug fixes, new features, or documentation improvements, feel free to submit an issue or a pull request.
If you modify or extend MeshFL in a derivative work intended for publication (such as a research paper or software tool), please cite and acknowledge the original MeshFL project and the original authors.
We also request that significant contributions to derivative works be recognized by including original authors as co-authors, where appropriate.
NVFlare: Federated learning framework.
MeshNet: Volumetric dilated convolutional neural network architecture for MRI segmentation.
MeshFL release V1.0.0 was funded by the NIH grant R01DA040487.
