3D U-net for volumetric segmentation of spinal cord MRI in degenerative disorders.

Abstract

The correct identification of spinal cord structures in magnetic resonance imaging (MRI) plays a vital role in identifying degenerative spondyloarthrosis, osteophytes, osteochondrosis, haemangioma, osteophytes and physiological lordosis malformations. Delineation, which is done physically, is time-consuming and can be varied. We will present a 3D U-Net-based deep learning model of volumetric segmentation of spinal cord MRI data in Digital Imaging and Communications in Medicine (.dcm) format in this research. The model takes advantage of three-dimensional convolutions to learn spatial dependencies in each of the axial, sagittal, and coronal planes. MRI volumes were pre-processed using normalisation, artefact removal, and volumetric resampling, resulting in a filtered dataset containing patients with multiple degenerative changes. The comparison was done with performance against expert annotations under the Dice similarity coefficient, Hausdorff distance, sensitivity, and specificity. Findings indicate that the 3D U-Net scores higher than the other models on dominant spinal cord structures. This research discusses the possibility of volumetric deep learning helping to optimise radiological processes and assist in precision medicine for spinal cord imaging.