Automatic Paraspinal Muscle Segmentation in Patients with Lumbar Pathology Using Deep Convolutional Neural Network

Abstract

Recent evidence suggests an association between low back pain (LBP) and changes in lumbar paraspinal muscle morphology and composition (i.e., fatty infiltration). Quantitative measurements of muscle cross-sectional areas (CSAs) from MRI scans are commonly used to examine the relationship between paraspinal muscle characters and different lumbar conditions. The current investigation primarily uses manual segmentation that is time-consuming, laborious, and can be inconsistent. However, no automatic MRI segmentation algorithms exist for pathological data, likely due to the complex paraspinal muscle anatomy and high variability in muscle composition among patients. We employed deep convolutional neural networks using U-Net+CRF-RNN with multi-data training to automatically segment paraspinal muscles from T2-weighted MRI axial slices at the L4-L5 and L5-S1 spinal levels and achieved averaged Dice score of 93.9 % and mean boundary distance of 1 mm. We also demonstrate the application using the segmentation results to reveal tissue characteristics of the muscles in relation to age and sex.