Automatic Segmentation of Spinal Cord Gray Matter from MR Images using a U-Net Architecture

Abstract

The human spinal cord is a highly organized and complex part of the central nervous system. In particular, GM in the spinal cord is associated with many neurological diseases such as multiple sclerosis (MS), amyotrophic lateral sclerosis (ALS), etc. In addition, the accurate determination of GM in the spinal cord by volume is very important for the diagnosis of spinal cord lesions and other neurological diseases at an early stage. Clinical symptoms/signs, cerebrospinal fluid examinations, evoked potentials, and magnetic resonance imaging (MRI) findings are used in the diagnosis of neurological diseases in the spinal cord region. However, since the spinal cord area does not have a definite geometric shape and is not flat along the back, artifacts often occur in MR scans obtained from the region and it is more difficult to determine the boundaries of the spinal cord area and to detect the lesions in this region. In this chapter, automatic segmentation of spinal cord GM on MR images using U-Net deep learning architecture is proposed. Spinal cord gray matter segmentation challenge (SCGMC) publicly available dataset is used in the study for experimental studies. In this dataset, the spinal cord GM region is successfully segmented using the U-Net architecture. In experimental studies, score of 0.83 is achieved for the dice similarity coefficient (DSC) in segmentation of GM. As a result, it has been confirmed that the spinal cord GM can be segmented with high accuracy with the U-Net architecture proposed in the study.