Light-weight neural network for intra-voxel structure analysis

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Diffusion-weighted MRI (dMRI) is crucial for studying brain connectivity and development.
• Analyzing intra-voxel structures in dMRI data presents significant challenges due to noise and complexity.
• Existing methods like constrained spherical deconvolution (CSD) have limitations in accuracy and spatial consistency.

Purpose of the Study:

• To introduce a novel neural network, the Local Neighborhood Neural Network (LNNN), for enhanced intra-voxel structure analysis in dMRI.
• To address the challenge of limited ground truth data using a self-supervised learning approach.
• To improve the accuracy and spatial consistency of dMRI analysis for brain imaging applications.

Main Methods:

• Developed the Local Neighborhood Neural Network (LNNN) architecture to leverage spatial correlations between neighboring voxels.
• Implemented a self-supervised learning strategy by generating synthetic voxel neighborhood signals for training.
• Trained the LNNN model on phantom and real brain imaging data.

Main Results:

• The LNNN method demonstrated superior performance compared to constrained spherical deconvolution (CSD) in quantitative and qualitative validations.
• Improved accuracy in angular error and volume fraction estimation was observed using phantom data.
• Enhanced spatial consistency in real brain images was achieved, outperforming CSD.

Conclusions:

• The Local Neighborhood Neural Network offers a robust and accurate method for intra-voxel structure analysis in diffusion-weighted MRI.
• The self-supervised approach effectively overcomes the limitations of ground truth data scarcity.
• This novel method shows significant potential for advancing brain connectivity and development research using neuroimaging.