Automatic segmentation of all lower limb muscles from high-resolution magnetic resonance imaging using a cascaded three-dimensional deep convolutional neural network

Abstract. High-resolution magnetic resonance imaging with fat suppression can obtain accurate anatomical information of all 35 lower limb muscles and individual segmentation can facilitate quantitative analysis. However, due to limited contrast and edge information, automatic segmentation of the muscles is very challenging, especially for athletes whose muscles are all well developed and more compact than the average population. Deep convolutional neural network (DCNN)-based segmentation methods showed great promise in many clinical applications, however, a direct adoption of DCNN to lower limb muscle segmentation is challenged by the large three-dimensional (3-D) image size and lack of the direct usage of muscle location information. We developed a cascaded 3-D DCNN model with the first step to localize each muscle using low-resolution images and the second step to segment it using cropped high-resolution images with individually trained networks. The workflow was optimized to account for different characteristics of each muscle for improved accuracy and reduced training and testing time. A testing augmentation technique was proposed to smooth the segmentation contours. The segmentation performance of 14 muscles was within interobserver variability and 21 were slightly worse than humans.