Methods

Model Architecture

ThighMRIseg uses SegResNet, a 3D encoder-decoder segmentation network from the MONAI framework.

3D MRI Input → SegResNet Encoder → Decoder → 11-class Segmentation Map

Key Features

• Residual connections for stable training
• 3D convolutions capturing volumetric context
• Multi-scale feature aggregation

Training Strategy

• Framework: MONAI Label
• Loss function: Dice + Cross-Entropy
• Data augmentation: Random flipping, rotation, intensity scaling
• Modalities: Trained jointly on IDEAL fat/water, T1, T2, and STIR sequences

Deployment

The model is served via MONAI Label, enabling:

• REST API for automated inference
• Integration with OHIF Viewer (web-based)
• Integration with 3D Slicer (desktop)
• Batch processing for research studies

Post-Processing

• Connected component analysis for noise removal
• Morphological operations for boundary refinement
• Per-muscle volume and cross-sectional area computation