Generalizable lower-limb muscle MRI segmentation and quantification on heterogeneous multisite datasets

Neuromuscular diseases are a heterogeneous group of disorders affecting muscles and peripheral nerves, leading to progressive muscle weakness and functional impairment. Muscle MRI facilitates the assessment of muscle pathology, but current analysis relies on time-consuming manual segmentation or subjective visual scoring. Validation across diverse patient populations and imaging protocols is lacking in existing automated methods. Deep learning segmentation methods were developed to quantify intramuscular fat infiltration and muscle volume across all lower limb muscles using heterogeneous multi-site data. Three convolutional neural network architectures (U-Net, U-Net + + , and Attention U-Net + +) were evaluated on a multi-site dataset comprising 27,858 slices from 797 muscle MRI scans across 376 patients, spanning 12 neuromuscular diseases and 12 international sites. Thirty-two individual muscles across pelvis, thigh, and lower leg regions were segmented. High accuracy was achieved by all architectures (DSC = 0.97). Leave-One-Site-Out experiments revealed strong generalisability across sites (average DSC = 0.96). Automated fat quantification placed 90.9% of muscles within one point of ground truth on standard visual scales and correlated strongly with quantitative fat fraction measurements (r = 0.995, p < 0.001). High correlation with ground truth was demonstrated by cross-sectional area predictions (r = 0.94, p < 0.001). Diverse imaging protocols were handled with minimal preprocessing. Accuracy equivalent to observer variability was achieved by automated segmentation, potentially eliminating the need for manual correction in large-scale applications.