qModeL: A plug-and-play model-based reconstruction for highly accelerated multi-shot diffusion MRI using learned...

Area of Science:

• Medical Imaging
• Neuroimaging
• Diffusion MRI

Background:

• Multi-shot Echo Planar Imaging (EPI) offers higher spatial resolution in diffusion MRI but requires long scan times.
• Accelerated multi-shot diffusion weighted (msDW) scans are crucial for advanced microstructure studies needing extensive q-space coverage.
• Existing joint k-q undersampling methods with compressed sensing struggle with multi-shell data reconstruction due to limitations of sparsity priors.

Purpose of the Study:

• To introduce a novel joint reconstruction method, qModeL, for highly undersampled multi-shot diffusion weighted (msDW) scans.
• To enable accurate reconstruction of diffusion MRI data from accelerated multi-shot acquisitions, facilitating advanced microstructure analysis.

Main Methods:

• Proposed qModeL reconstruction integrates model-based iterative techniques with machine learning, extending plug-and-play algorithms.
• Deep learning is used to pre-learn the signal manifold in the diffusion measurement space, providing voxel-wise q-dimension regularization.
• The framework incorporates plug-and-play total variation denoising for spatial dimension regularization and is validated on single-shell and multi-shell msDW data.

Main Results:

• qModeL successfully reconstructs diffusion weighted images (DWIs) from 8-fold accelerated msDW acquisitions with less than 5% error for both single-shell and multi-shell data.
• Advanced microstructural analyses performed using the qModeL-reconstructed data demonstrate reasonable accuracy.

Conclusions:

• qModeL facilitates joint recovery of highly accelerated multi-shot diffusion MRI data using learned, biophysically-driven priors.
• This approach supports the utilization of accelerated multi-shot imaging for multi-shell sampling and in-depth microstructure studies.