Isotropic Reconstruction of MR Images Using 3D Patch-Based Self-Similarity Learning

Area of Science:

• Medical Imaging
• Image Reconstruction
• Magnetic Resonance Imaging (MRI)

Background:

• Isotropic 3D MRI acquisition is challenging, particularly in cardiac applications, due to hardware and time constraints leading to low resolution.
• Current super-resolution (SR) methods for isotropic reconstruction often rely on local regularization, which can struggle with noise amplification.
• Existing techniques result in anisotropic 3D volumes with poor image quality in the through-plane direction.

Purpose of the Study:

• To develop a novel isotropic 3D reconstruction scheme for MRI.
• To improve image quality in low-resolution 3D acquisitions, especially for cardiac MRI.
• To integrate non-local and self-similarity information for robust image reconstruction.

Main Methods:

• A novel isotropic 3D reconstruction scheme integrating non-local and self-similarity information from 3D patch neighborhoods.
• Grouping 3D patches with similar structures to enforce natural sparsity and low-rank properties of MR images.
• Utilizing an Augmented Lagrangian formulation to decompose the optimization into low-rank volume denoising and SR reconstruction.

Main Results:

• The proposed joint SR and self-similarity learning framework demonstrated superior performance compared to state-of-the-art methods.
• Experimental results in simulations, brain imaging, and clinical cardiac MRI validated the effectiveness of the proposed method.
• The method achieved robust image reconstruction with high signal-to-noise ratio efficiency.

Conclusions:

• The novel isotropic 3D reconstruction scheme effectively enhances image quality in MRI.
• The integration of self-similarity learning provides a robust approach for super-resolution reconstruction.
• This technique holds significant potential for cardiac MRI applications, including myocardial infarction scar assessment.