Self-prior image-guided MRI reconstruction with dictionary learning.

Area of Science:

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

Background:

• Accelerated MRI acquisition is crucial for reducing scan times and improving patient comfort.
• Traditional methods often struggle with image quality at high acceleration rates.
• Prior image-guided reconstruction typically requires separate reference scans, limiting its applicability.

Purpose of the Study:

• To develop a novel method, SPIDLE, for improving MRI reconstruction performance at high acceleration rates.
• To enable prior image-guided reconstruction without the need for additional MRI scans.
• To enhance the quality of highly undersampled MRI data.

Main Methods:

• SPIDLE integrates self-prior image constraint, compressed sensing (CS), and dictionary learning (DL).
• A preliminary reconstruction (self-prior image) is derived from the target image's undersampled data.
• CS is applied to the residual error between the target and self-prior images, leveraging sparsity.
• The split Bregman method is used for efficient optimization and fast convergence.

Main Results:

• SPIDLE demonstrated superior performance compared to state-of-the-art methods across various in vivo datasets and undersampling schemes.
• The method achieved fewer reconstruction errors and showed robustness to initialization.
• Reconstruction quality was significantly improved, especially under high undersampling conditions.

Conclusions:

• SPIDLE expands the utility of prior image-guided MRI reconstruction, particularly in scenarios where external prior scans are impractical.
• The method offers substantial improvements in reconstruction quality for highly undersampled MRI.
• SPIDLE shows promise for dynamic MRI and multimodal imaging, including CT and CT-MRI.