TrGANet: Transforming 3T to 7T dMRI using Trapezoidal Rule and Graph based Attention Modules

Area of Science:

• Neuroimaging
• Medical Physics
• Computer Vision

Background:

• Diffusion MRI (dMRI) is crucial for brain white matter analysis, aiding in disease diagnosis and surgical planning.
• Higher angular resolution diffusion imaging (HARDI) and higher magnetic strengths (e.g., 7T) offer improved tractography and anatomical detail.
• High-field MRI scanners (7T) are expensive, limiting their widespread clinical adoption.

Purpose of the Study:

• To develop a novel Convolutional Neural Network (CNN) architecture for transforming 3T dMRI data to 7T quality.
• To reconstruct multi-shell multi-tissue fiber orientation distribution functions (MSMT fODF) at 7T from single-shell 3T data.
• To enhance the accessibility and quality of diffusion MRI for clinical applications.

Main Methods:

• Proposed a novel CNN architecture incorporating an ODE solver with the Trapezoidal rule.
• Integrated a graph-based attention layer into the CNN architecture.
• Utilized L1 and total variation loss functions for model training and validation.

Main Results:

• Successfully transformed 3T dMRI data to achieve 7T-like quality.
• Reconstructed MSMT fODF at 7T from 3T single-shell data.
• Validated the model quantitatively and qualitatively on the Human Connectome Project (HCP) dataset.

Conclusions:

• The proposed CNN architecture effectively simulates 7T dMRI quality from lower-field 3T data.
• This method can significantly improve the diagnostic capabilities and surgical planning utility of dMRI.
• Enhancing dMRI quality from 3T to 7T can democratize access to advanced neuroimaging techniques.