Tensor-decomposition regularized learning for fast and high-fidelity multi-parametric microstructural MR imaging.

Area of Science:

• Neuroimaging
• Medical Physics
• Artificial Intelligence

Background:

• Deep learning (DL) shows promise for analyzing diffusion-weighted MR images to understand brain microstructures.
• Current methods struggle with efficient and accurate estimation of multiple microstructural parameters due to isolated modeling and dense sampling needs.

Purpose of the Study:

• To develop a unified DL framework for fast, high-fidelity estimation of multiple microstructural parameters from various diffusion models.
• To achieve this using sparsely sampled q-space data.

Main Methods:

• Proposed DeepMpMRI, an efficient DL framework with a novel tensor-decomposition-based regularizer.
• Incorporated a Nesterov-based adaptive learning algorithm to optimize regularization parameters dynamically.

Main Results:

• DeepMpMRI outperformed five state-of-the-art methods on HCP and Alzheimer's datasets.
• Achieved 4.5-15x acceleration in estimating DKI- and NODDI-derived parameter maps compared to dense sampling.

Conclusions:

• DeepMpMRI facilitates accurate and robust multi-parametric microstructural imaging under sparse sampling.
• Demonstrates significant potential for clinical translation in diffusion MRI-based tissue characterization.