Re-Visible Dual-Domain Self-Supervised Deep Unfolding Network for MRI Reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Magnetic Resonance Imaging (MRI) acquisition is time-consuming, limiting its clinical utility.
• Existing deep learning acceleration methods require extensive fully-sampled data for supervised training.
• Self-supervised methods face challenges with input distribution shifts and insufficient incorporation of image priors.

Purpose of the Study:

• To develop a novel self-supervised deep unfolding network for accelerated MRI acquisition using only under-sampled datasets.
• To address the input distribution shift issue inherent in current self-supervised learning approaches.
• To enhance MRI reconstruction performance by effectively integrating imaging physics and image priors.

Main Methods:

• Proposed a re-visible dual-domain self-supervised deep unfolding network.
• Incorporated a re-visible dual-domain loss to utilize all under-sampled k-space data, mitigating training-inference distribution shifts.
• Developed a Deep Unfolding Network based on Chambolle and Pock Proximal Point Algorithm (DUN-CP-PPA) with a Spatial-Frequency Feature Extraction (SFFE) block for end-to-end reconstruction.

Main Results:

• The proposed method effectively mitigates the input distribution shift by utilizing all available under-sampled k-space data.
• The integration of imaging physics and comprehensive image priors through the SFFE block significantly enhances reconstruction performance.
• Experimental results on single-coil and multi-coil datasets demonstrate superior reconstruction quality and generalization capability compared to state-of-the-art methods.

Conclusions:

• The novel self-supervised deep unfolding network offers a promising solution for accelerated MRI acquisition without requiring fully-sampled data.
• The re-visible dual-domain approach and DUN-CP-PPA architecture effectively address limitations of previous self-supervised methods.
• The method achieves state-of-the-art performance, paving the way for more efficient and widely applicable MRI techniques.