A Unified Learning Model for Estimating Fiber Orientation Distribution Functions on Heterogeneous Multi-shell...

Area of Science:

• Neuroimaging
• Biomedical Engineering
• Computational Neuroscience

Background:

• Diffusion-weighted MRI (DW-MRI) measures water diffusion in biological tissues, crucial for understanding microstructure.
• Recent advances focus on radial b-value dependence for improved tissue classification and micro-architecture estimation.
• Existing deep learning methods often require multi-stage strategies and rely on intermediate representations.

Purpose of the Study:

• To develop a unified, single-stage deep learning network for efficient fiber orientation distribution function (fODF) estimation.
• To enable accurate fODF estimation from heterogeneous multi-shell DW-MRI sequences.
• To compare the performance of the proposed single-stage method against traditional multi-stage approaches.

Main Methods:

• A novel single-stage spherical convolutional neural network was developed.
• The network was trained and validated using Human Connectome Project (HCP) young adult test-retest scan data.
• Performance was evaluated using heterogeneous multi-shell, shell-dropoff, and single-shell DW-MRI sequences.

Main Results:

• The proposed single-stage network demonstrated efficient and accurate fODF estimation.
• The method outperformed prior multi-stage deep learning approaches.
• Robust performance was observed across various DW-MRI sequence types, including single-shell data.

Conclusions:

• The developed unified dynamic network offers a more efficient and effective approach for fODF estimation in DW-MRI.
• This single-stage method simplifies the deep learning pipeline for microstructure imaging.
• The findings suggest potential for improved diagnostic and research applications in neuroscience.