Undersampled single-shell to MSMT fODF reconstruction using CNN-based ODE solver

Area of Science:

• Neuroimaging
• Diffusion MRI (dMRI)
• Computational Neuroscience

Background:

• Diffusion MRI (dMRI) is crucial for non-invasively studying white matter (WM) microstructure and fiber tracts.
• Advanced techniques like Multi-Shell Multi-Tissue fiber orientation distribution function (MSMT fODF) offer precise fiber directionality but require long scan times.
• Current clinical dMRI scanners often acquire limited single-shell data due to SNR and artifact constraints.

Purpose of the Study:

• To develop a method for reconstructing MSMT fODF from under-sampled or fully sampled single-shell dMRI data.
• To overcome the limitations of long scanning times associated with multi-shell dMRI acquisition.
• To enable more accurate white matter tractography and microstructural analysis in clinical settings.

Main Methods:

• Proposed a Convolutional Neural Network (CNN)-based ordinary differential equations solver.
• The architecture incorporates CNN-based Adams-Bashforth and Runge-Kutta modules.
• Utilized L1 and total variation loss functions for model training and optimization.

Main Results:

• Successfully reconstructed MSMT fODF, fiber tracts, and structural connectivity using the HCP dataset.
• Achieved high angular correlation coefficients for white matter and the full brain, demonstrating network utility.
• Validated network robustness against varying signal-to-noise ratios (SNR).

Conclusions:

• The proposed CNN model accurately predicts MSMT fODF from single-shell dMRI volumes.
• This approach effectively addresses the challenge of acquiring multi-shell data for detailed brain microstructure analysis.
• The method holds promise for improving the clinical applicability of advanced dMRI techniques.