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

White matter tract segmentation is critical for mapping brain connectivity in both clinical and research settings. Recent deep learning methods have enabled direct voxel-wise segmentation from diffusion MRI (dMRI), bypassing tractography. However, most approaches rely on a limited number of peaks extracted from the fiber orientation distribution function (fODF) at each voxel, which discards important orientation information, particularly in problematic regions with complex fiber configurations such as crossing fibers and bottlenecks.

In this work, we introduce FODSeg, a voxel-based segmentation method that utilizes the complete fODF representation for each voxel, capturing the full angular structure of white matter orientation. Additionally, we reformulate tract segmentation as a singleclass problem, training one model per tract to reduce label conflicts inherent in multi-class approaches. This combination allows FODSeg to better distinguish tracts with similar local orientations and improves robustness in regions with structural ambiguity. We evaluate FODSeg on the Human Connectome Project dataset across all 72 white matter tracts using six segmentation accuracy metrics.

FODSeg achieves higher Dice scores and lower volumetric overreach values in 70% of the tracts while maintaining high specificity. Our results demonstrate the superior performance of FODSeg over existing segmentation approaches. Notably, our method shows significant improvements in anatomically challenging bottleneck regions, reducing false positives and improving tract-specific precision.

Overall, FODSeg advances white matter tract segmentation by leveraging the full richness of the fODF signal while improving accuracy, specificity, and anatomical consistency.