Deep Learning-Based Auto-Segmentation of Spinal Cord Internal Structure of Diffusion Tensor Imaging in Cervical...

Area of Science:

• Neuroimaging
• Medical image analysis
• Spinal cord imaging

Background:

• Cervical spondylotic myelopathy (CSM) is a chronic spinal cord disorder.
• Diffusion tensor imaging (DTI) provides valuable insights into spinal cord status for CSM diagnosis and prognosis.
• Manual extraction of DTI features from regions of interest (ROIs) is labor-intensive and time-consuming.

Purpose of the Study:

• To develop and validate an automated segmentation model for DTI-based ROIs in the cervical spinal cord.
• To assess the efficiency and accuracy of the automated model compared to manual segmentation.
• To evaluate the potential of automated segmentation for detailed quantification of cervical spinal cord status.

Main Methods:

• Analysis of 1159 cervical spinal cord slices from 89 CSM patients.
• Calculation of fractional anisotropy (FA) maps from DTI data.
• Training a UNet model with heatmap distance loss for auto-segmentation of eight ROIs (lateral, dorsal, ventral columns, and gray matter).

Main Results:

• The UNet model achieved mean Dice coefficients ranging from 0.54 to 0.69 across different ROIs and sides.
• Segmented ROI-based mean FA values strongly correlated with manually drawn values.
• Low mean absolute errors (0.07-0.11) were observed between automated and manual segmentation values.

Conclusions:

• The proposed UNet-based segmentation model automates DTI feature extraction from cervical spinal cord ROIs.
• This automated approach offers a more detailed and efficient method for spinal cord segmentation.
• The model shows potential for improved quantification of spinal cord status in CSM patients, aiding diagnosis and prognosis.