M-MRI: A Manifold-based Framework to Highly Accelerated Dynamic Magnetic Resonance Imaging

Area of Science:

• Medical Imaging
• Signal Processing
• Computational Geometry

Background:

• Dynamic magnetic resonance (dMR) images are high-dimensional signals often residing on low-dimensional manifolds.
• Current dynamic magnetic resonance imaging (dMRI) reconstruction methods frequently utilize low-rank and sparsity priors.
• Reconstructing dMRI from highly undersampled k-space data remains a significant challenge.

Purpose of the Study:

• To propose a novel manifold-based framework, M-MRI, for dMRI reconstruction.
• To model dMRI images as points on a smooth manifold and learn its geometry.
• To develop regularization loss functions based on learned manifold geometry for improved reconstruction.

Main Methods:

• The M-MRI framework models dMRI images as points on a learned manifold.
• Navigator signals are used to learn the underlying manifold geometry.
• Low-dimensional embeddings preserving manifold geometry are computed.
• Two regularization loss functions are derived from the learned manifold geometry.

Main Results:

• The M-MRI framework demonstrates effective dMRI reconstruction from highly undersampled k-space data.
• Validation was performed using extensive numerical tests on both phantom and in-vivo datasets.
• The proposed manifold-based approach shows promise for advanced dMRI reconstruction.

Conclusions:

• The M-MRI framework offers a novel approach to dMRI reconstruction by utilizing manifold learning.
• The method effectively reconstructs dynamic MR images from undersampled k-space data.
• This manifold-based strategy has the potential to enhance dMRI acquisition and analysis.