Fiber HGNN: Heterogeneous Graph Neural Network for Fiber Tract Segmentation.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Artificial Intelligence

Background:

• Accurate fiber tract segmentation is vital for brain function interpretation and surgical planning.
• Current methods struggle to integrate heterogeneous information like streamline shape, point position, and anatomical priors.
• Existing approaches often rely on either cortical-parcellation-based or fiber clustering methods.

Purpose of the Study:

• To propose Fiber HGNN, a novel heterogeneous graph neural network (HGNN) for accurate fiber tract segmentation.
• To explicitly model and integrate heterogeneous information from brain fibers.
• To improve upon existing methods by simultaneously leveraging streamline shape, local geometry, and anatomical priors.

Main Methods:

• Constructed a heterogeneous graph with streamline, fiber keypoint, and anatomical region nodes.
• Developed a Metapath-guided Heterogeneous Information Aggregation (MHIA) network to leverage implicit anatomical connectivity.
• Decomposed the heterogeneous graph into anatomical subgraphs for each streamline, aggregating information via metapath-linked nodes.

Main Results:

• Fiber HGNN effectively integrates complementary information from streamline shape, local geometry, and anatomical priors.
• The MHIA network enhances feature representation learning by considering anatomical context.
• Experimental results on HCP105 and TractoInferno datasets show significant performance improvements over state-of-the-art methods.

Conclusions:

• Fiber HGNN offers a powerful new framework for accurate fiber tract segmentation by integrating heterogeneous information.
• The proposed method advances the state-of-the-art in neuroimaging analysis for clinical applications.
• The developed approach facilitates more discriminative feature representations for improved tract segmentation accuracy.