A Multimodal Deep Learning Approach for White Matter Shape Prediction in Diffusion MRI Tractography.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• White matter tractography shape measures offer insights into brain anatomy and disease.
• Conventional methods for shape measure computation are computationally intensive and slow for large datasets.

Purpose of the Study:

• To introduce Tract2Shape, a novel multimodal deep learning framework for efficient and accurate prediction of white matter shape measures.
• To address the computational limitations of traditional voxel-based shape analysis.

Main Methods:

• Developed a multimodal deep learning framework (Tract2Shape) integrating geometric streamline and scalar data.
• Utilized a Siamese architecture with dual-encoder networks and dimensionality reduction (PCA).
• Trained and evaluated on Human Connectome Project (HCP-YA) and Parkinson's Progression Markers Initiative (PPMI) datasets.

Main Results:

• Tract2Shape outperformed state-of-the-art models on the HCP-YA dataset, achieving high accuracy (Pearson's r) and low error (nMSE).
• Demonstrated strong generalizability on the unseen PPMI dataset.
• Achieved a 99.2% improvement in computational efficiency compared to traditional methods.

Conclusions:

• Tract2Shape enables fast, accurate, and generalizable prediction of white matter shape measures.
• The framework supports scalable analysis for large-scale neuroimaging datasets.
• Paves the way for advanced large-scale white matter shape analysis in neuroscience research.