FODSeg: a deep learning framework for tract-specific white matter segmentation from full angular distributions.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Medical Image Analysis

Background:

• White matter tract segmentation is crucial for understanding brain connectivity.
• Current deep learning methods using limited fiber orientation distribution function (fODF) peaks struggle with complex regions like crossing fibers and bottlenecks.
• Existing approaches often discard valuable orientation information, impacting segmentation accuracy.

Purpose of the Study:

• To introduce FODSeg, a novel voxel-based segmentation method for white matter tracts.
• To leverage the complete fODF representation for improved angular structure capture.
• To enhance segmentation accuracy and robustness in challenging neuroanatomical regions.

Main Methods:

• FODSeg utilizes the full fODF representation at each voxel, capturing comprehensive orientation information.
• Tract segmentation is reformulated as a single-class problem, training one model per tract to minimize label conflicts.
• The method was evaluated on the Human Connectome Project dataset across 72 white matter tracts.

Main Results:

• FODSeg achieved higher Dice scores and lower volumetric overreach in 70% of tracts compared to existing methods.
• The approach demonstrated high specificity and significant improvements in anatomically challenging bottleneck regions.
• Reduced false positives and enhanced tract-specific precision were observed.

Conclusions:

• FODSeg represents an advancement in white matter tract segmentation by utilizing the complete fODF signal.
• The method improves accuracy, specificity, and anatomical consistency in brain connectivity mapping.
• FODSeg offers a more robust solution for segmenting complex white matter structures.