MRI image segmentation of the major lower leg muscles using deep learning: application in biomechanical analysis.

objectiveMuscle volume is one of the major morphological parameters determining force and joint torque production capacity. An accurate, automatic muscle segmentation method, using imaging modalities such as Magnetic Resonance Image (MRI), is critical for precise volume measurements thus enhancing the utility of MRI-based quantitative muscle evaluation.

methodsWe developed EdgeUNETR++, an encoder-decoder model based on a transformer architecture that incorporates an edge enhancement module. Our proposed model was applied to on an MRI dataset of major lower leg muscles from both able-bodied and post-stroke participants to extract the muscle volume information. To further assess its potential for use in biomechanical analysis, we evaluated the correlation between the segmented muscle volumes and optimal ankle joint torque.

resultsEdgeUNETR + + achieved segmentation performance comparable to the baseline model, with modest improvements in certain muscles and in boundary-focused metrics. Across all segmentation scenarios, muscle volumes derived from both models were close to the ground truth and exhibited a similar positive trend in their relationship with optimal ankle torque across the four major muscles (i.e., soleus, lateral gastrocnemius, medial gastrocnemius, and tibialis anterior).

conclusionIn conclusion, EdgeUNETR + + may serve as an automated and practical tool for lower leg muscle segmentation and volume estimation from MRI, supporting investigations of muscle morphology and related biomechanical analyses.