Multi-Shell D-MRI Reconstruction via Residual Learning utilizing Encoder-Decoder Network with Attention (MSR-Net).

Area of Science:

• Medical Imaging
• Neuroscience
• Machine Learning

Background:

• Diffusion MRI with High Angular Resolution Diffusion Imaging (HARDI) offers accurate fiber orientation but faces limitations with single b-value acquisitions.
• Single-shell HARDI can lead to biased and noisy estimations of fiber Orientation Distribution Functions (ODFs) due to inability to determine volume fractions.
• Multi-shell HARDI overcomes these limitations but requires longer scanning times, hindering clinical applicability.

Purpose of the Study:

• To develop a novel deep learning architecture, MSR-Net, for reconstructing diffusion MRI volumes at one b-value using data from another.
• To enable more accurate fiber orientation analysis in diffusion MRI without extending acquisition time.

Main Methods:

• Proposed a novel deep learning architecture, MSR-Net, featuring an encoder-decoder structure with attention and feature modules.
• Learned transformations in the space of spherical harmonic coefficients for reconstructing diffusion MRI data between b = 2000 s/mm² and b = 1000 s/mm².
• Utilized L2 and Content loss functions for network optimization and performance enhancement, validated on the Human Connectome Project (HCP) dataset.

Main Results:

• Demonstrated the capability of MSR-Net to reconstruct diffusion MRI volumes between different b-values (2000 s/mm² and 1000 s/mm²).
• Achieved standard qualitative and quantitative performance measures, indicating successful network training and validation.
• Showcased the potential for improved fiber orientation estimation using the reconstructed data.

Conclusions:

• MSR-Net offers a promising solution for enhancing diffusion MRI analysis by enabling accurate reconstruction between b-values.
• This deep learning approach can potentially mitigate the trade-off between data quality and scanning time in clinical settings.
• The method facilitates more robust fiber orientation estimation, crucial for understanding brain connectivity.