Deep Learning Based Automatic DTI Analysis For Cervical Myelopathy

Cervical myelopathy (CM) is a prevalent degenerative spinal disorder that primarily affects the elderly population. Diffusion Tensor Imaging (DTI) holds promise for precise diagnosis of CM by quantitatively investigating the ultra-structure. However, the manual column-specific ROI definition method is time-consuming and prone to measurement variations due to unstandardized ROI definitions and operator bias. In this study, we propose an automatic DTI segmentation method using deep learning techniques to delineate ROIs in the cervical spinal cord. We developed an auto-segmentation tool based on the nnU-Net framework and recruited 105 patients for training and evaluation. We collected several original DTI images and several other DTI metrics maps, including B0, E0, E1, E2, and FA images, to construct several datasets. The nnU-Net-based segmentation model achieved superior results on a combined dataset comprising B0, E0, E1, E2, and FA images with a dice similarity coefficient of 0.55 on the test dataset. Furthermore, our nnU-Net model demonstrated improved performance across different ROIs with a maximum dice similarity coefficient value of 0.65 observed on the left side of the lateral column and a minimum value of 0.47 in the left ventral column. This study highlights the feasibility of utilizing deep learning models for more detailed segmentation of the spinal cord in DTI images.