Towards Accurate Microstructure Estimation via 3D Hybrid Graph Transformer.

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Deep learning methods are increasingly used for microstructure estimation in undersampled diffusion MRI (dMRI) data.
• The hybrid graph transformer (HGT) integrates k-space graph learning and q-space transformer learning but neglects 3D spatial information by using 2D slices.

Purpose of the Study:

• To propose an advanced microstructure estimation model, 3D hybrid graph transformer (3D-HGT), that leverages full 3D spatial and angular information.
• To develop an efficient k-space learning model to address the computational burden of 3D k-space learning.
• To introduce a 3D q-space learning module utilizing transformer architecture.

Main Methods:

• Development of the 3D hybrid graph transformer (3D-HGT) model.
• Implementation of an efficient k-space learning component using simplified graph neural networks.
• Integration of a 3D q-space learning module based on transformer architecture.
• Validation using Human Connectome Project dMRI data.

Main Results:

• The proposed 3D-HGT model effectively utilizes 3D spatial and angular information for microstructure estimation.
• Experiments demonstrated that 3D-HGT significantly outperforms existing state-of-the-art methods, including HGT.
• Both quantitative and qualitative evaluations confirmed the superior performance of 3D-HGT.

Conclusions:

• 3D-HGT represents a significant advancement in dMRI microstructure estimation by incorporating 3D spatial context.
• The model offers improved accuracy and performance compared to previous methods, particularly those relying on 2D data.
• This work highlights the potential of 3D deep learning approaches for enhanced neuroimaging analysis.