Artificial intelligence in corticospinal tract segmentation using constrained spherical deconvolution.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• Cerebral white matter tractography is crucial for neurosurgical planning and diagnosing neurological diseases.
• Constrained spherical deconvolution (CSD) offers efficient and plausible segmentations for tractography.
• Comparing CSD techniques is essential for optimizing corticospinal tract (CST) segmentation.

Purpose of the Study:

• To compare two CSD techniques for segmenting the corticospinal tract (CST).
• To evaluate the similarity and consistency of CST segmentation using region-based and automatic (TractSeg) approaches.

Main Methods:

• Utilized 40 diffusion-weighted images (DWIs) from the Human Connectome Project (HCP) and 12 clinical DWIs.
• Performed tractography using a regions of interest-based approach and the TractSeg neural network.
• Quantified segmentation overlap using the Dice similarity coefficient and assessed consistency with intraclass correlation coefficients.

Main Results:

• Low similarity was found between the two CST segmentation methods (Dice index: HCP 0.479, Clinical 0.404).
• Both techniques demonstrated high consistency in sequential measurements (ICC > 0.995).
• Significant differences were observed in volume, fractional anisotropy (FA), and mean diffusivity (MD) between the methods.

Conclusions:

• Both CSD techniques provide consistent CST segmentation, with the automatic approach showing higher overall consistency.
• Moderate similarity and metric differences highlight distinct characteristics of each segmentation method.
• The findings underscore the importance of method selection in CSD-based tractography for clinical and research applications.