Hybrid Attentive Unet for Segmentation of Lower Leg Muscles and Bones From MRI Scans For Musculoskeletal Research

Musculoskeletal research such as studies of muscle growth in children with cerebral palsy (CP) often requires segmentation of muscles from magnetic resonance imaging (MRI) scans. This process has recently been automated by deep neural networks due to the costly and subjective nature of manual labelling. Deep neural networks typically perform well but, on average, tend to perform worse than human raters. Furthermore, deep neural networks need to generalize to scans of children with CP, which look different from scans of typically developing children because of differences in muscle size and composition, and typically constitute only a small portion of training data. To tackle those issues, we propose a novel end-to-end attention-based hybrid network that learns to segment musculoskeletal structures with a mixture of inter- and intra-slice features. The proposed network statistically significantly outperforms its contenders by a substantial margin and demonstrates robust generalization capabilities on scans of children with and without CP.