TractGeoNet: A geometric deep learning framework for pointwise analysis of tract microstructure to predict language...

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Diffusion magnetic resonance imaging (dMRI) tractography is crucial for mapping white matter pathways.
• Traditional dMRI tractometry methods often average or bin streamline data, potentially losing granular information.
• Relating white matter microstructure to cognitive functions like language remains a key challenge.

Purpose of the Study:

• To introduce TractGeoNet, a geometric deep learning framework for regression using dMRI tractography and pointwise tissue microstructure.
• To develop a novel Paired-Siamese Regression loss function for improved prediction accuracy.
• To create a Critical Region Localization algorithm for identifying predictive anatomical areas within fiber tracts.

Main Methods:

• Utilized a point cloud representation to directly process pointwise microstructure and positional information from dMRI tractography.
• Implemented a Paired-Siamese Regression loss to focus on relative differences in prediction scores.
• Applied a Critical Region Localization algorithm to pinpoint highly predictive anatomical regions.
• Evaluated on 806 subjects from the Human Connectome Project Young Adult dataset, predicting language performance from 20 association white matter tracts.

Main Results:

• TractGeoNet demonstrated superior prediction performance compared to existing regression models for cognitive performance based on neuroimaging features.
• The left arcuate fasciculus was identified as the most predictive tract for language performance.
• Critical regions within tracts, including temporal and frontal areas, were localized as consistently predictive of language abilities.

Conclusions:

• Geometric deep learning, via TractGeoNet, offers a powerful approach to analyze white matter fiber tracts and their relationship to human traits.
• The framework effectively utilizes pointwise dMRI data and identifies key brain regions influencing language performance.
• TractGeoNet advances the study of brain structure-function relationships, particularly for cognitive assessments.