Hybrid Graph Transformer for Tissue Microstructure Estimation with Undersampled Diffusion MRI Data.

Area of Science:

• Medical Imaging
• Neuroscience
• Computational Biology

Background:

• Diffusion MRI (dMRI) models require dense data for accurate tissue microstructure analysis.
• Sparse dMRI data limits the reliability of current microstructural modeling techniques.
• Existing deep learning methods often ignore diffusion wavevector space geometry or spatial neighborhood information.

Purpose of the Study:

• To develop a deep learning approach for predicting high-quality diffusion microstructural indices from sparsely sampled dMRI data.
• To explicitly incorporate the geometric structure of the diffusion wavevector space ($$-space) into the model.
• To leverage comprehensive spatial information in the physical coordinate space ($$-space).

Main Methods:

• Proposed a hybrid graph transformer (HGT) model.
• Utilized a graph neural network (GNN) to capture $$-space geometric structure.
• Implemented a novel residual dense transformer (RDT) for enhanced spatial information utilization.
• Trained and validated the model using data from the Human Connectome Project (HCP).

Main Results:

• The HGT model significantly improved the quality of microstructural estimations compared to existing methods.
• Explicitly considering $$-space geometry and leveraging spatial information led to superior performance.
• Demonstrated the effectiveness of the residual dense transformer for model training and accuracy.

Conclusions:

• The proposed hybrid graph transformer (HGT) offers a powerful solution for accurate diffusion microstructural analysis with sparse dMRI data.
• This method advances the field of diffusion MRI by enabling reliable microstructural index prediction even with limited sampling.
• The findings have significant implications for future research in neuroimaging and understanding tissue microstructure.