Self-Supervised Learning for Three-Dimensional Magnetic Resonance Imaging Reconstruction with Spatial Depth...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Accelerating Magnetic Resonance Imaging (MRI) acquisition is vital for clinical efficiency.
• Current deep learning MRI reconstruction methods often require fully sampled reference data, which are impractical to acquire.
• There is a need for self-supervised techniques that eliminate the dependency on reference scans.

Purpose of the Study:

• To propose a novel self-supervised deep learning method for MRI image reconstruction from undersampled k-space data.
• To eliminate the need for fully sampled reference data in MRI reconstruction.
• To improve the efficiency and feasibility of accelerated MRI acquisition.

Main Methods:

• A self-supervised framework utilizing two parallel reconstruction networks.
• Implementation of a spatial-depth attention mechanism for enhanced feature fusion in 3D data.
• Definition of a k-space reconstruction loss and a differential loss for network consistency.

Main Results:

• The proposed self-supervised method achieved comparable performance (PSNR/SSIM) to supervised methods.
• Effective reconstruction was demonstrated at 4x and 8x acceleration factors.
• The method successfully reconstructed images from undersampled k-space data without reference scans.

Conclusions:

• Self-supervised learning offers a viable alternative to supervised methods for accelerated MRI reconstruction.
• The developed spatial-depth attention mechanism and loss functions are effective for k-space data.
• This approach significantly advances the potential for rapid and efficient MRI acquisition.