Description:
We are excited to announce the first release of MeshDist NVFlare v1.0.0, a decentralized learning framework for 3D brain MRI segmentation using the MeshNet model. This release marks an important milestone, enabling distributed training with federated learning principles powered by NVFlare.
Key Features:
- Decentralized Training Framework:
Seamlessly integrates MeshNet with NVFlare for federated learning.
Supports multiple sites training independently while aggregating updates to improve global performance.
- MRI Data Handling:
Efficiently loads and processes 3D MRI data using SQLite-backed datasets. Includes Scanloader for dynamic dataset partitioning into training, validation, and testing subsets.
- Advanced Model Training:
OneCycleLR scheduler for optimal learning rate adjustment. Gradient accumulation for federated model optimization.
- Custom Metrics and Logging:
Supports Dice score calculation for segmentation performance evaluation. Integrated logging for detailed insights at each site, including training progress and model aggregation steps.
- NIfTI Output Support:
Generates 3D NIfTI files for model predictions, facilitating detailed visualization and analysis.
-Federated Learning Compatibility:
Fully compatible with NVFlare’s simulator for testing and development. Supports decentralized weight sharing and gradient aggregation workflows.
We welcome your feedback to help shape our priorities for MeshDist NVFlare. Looking forward to your contribution!