DeepBundle: Fiber Bundle Parcellation with Graph Convolution Neural Networks.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Accurate parcellation of whole-brain tractography streamlines is crucial for analyzing white matter microstructure.
• Current methods often fail due to difficulties in registering individual tractograms with atlases, stemming from significant inter-individual anatomical variations.

Purpose of the Study:

• To introduce DeepBundle, a novel deep learning approach for registration-free fiber parcellation.
• To overcome the limitations of traditional atlas-based registration in tract-based analysis.

Main Methods:

• Utilized graph convolution neural networks (GCNNs) to extract geometric features from individual fiber tracts.
• Developed a deep learning framework (DeepBundle) that predicts parcellation labels directly from these features, eliminating the need for registration.

Main Results:

• DeepBundle demonstrated effective registration-free fiber parcellation.
• The method successfully leveraged extracted geometric features for accurate tract segmentation.
• Evaluated using Human Connectome Project data, confirming the approach's advantages.

Conclusions:

• DeepBundle provides a robust alternative to traditional registration-dependent methods for fiber parcellation.
• Geometric features inherent to fiber tracts are sufficient for accurate parcellation.
• This deep learning approach advances tract-based analysis of brain white matter.