Self-Supervised Deep Unrolled Reconstruction Using Regularization by Denoising.

Area of Science:

• Medical Imaging
• Deep Learning
• Image Reconstruction

Background:

• Deep learning excels in computer vision and is increasingly applied to MRI reconstruction.
• Integrating deep learning with model-based optimization offers advantages but requires substantial labeled data.
• Data scarcity is a significant challenge for high-quality MRI reconstruction in many applications.

Purpose of the Study:

• To develop a novel, interpretable self-supervised learning method for MRI reconstruction.
• To reduce the dependency on large labeled datasets for high-quality MR image reconstruction.
• To enhance the performance of Noise2Noise methods in MRI reconstruction by incorporating imaging physics priors.

Main Methods:

• Proposed DURED-Net, combining a self-supervised denoising network with a plug-and-play method.
• Utilized Regularization by Denoising (RED) to leverage the denoising network for MRI reconstruction.
• Integrated explicit priors derived from imaging physics into the reconstruction process.

Main Results:

• DURED-Net achieves high reconstruction quality with a reduced amount of training data.
• The method demonstrates superior performance compared to state-of-the-art Noise2Noise approaches in MRI reconstruction.
• The integration of imaging physics priors enhances the effectiveness of deep learning-based reconstruction.

Conclusions:

• DURED-Net offers an effective solution for interpretable self-supervised MRI reconstruction.
• The method addresses the challenge of limited labeled data in MRI applications.
• This approach advances the field of deep learning in medical imaging by improving reconstruction efficiency and quality.