MRI RECOVERY WITH A SELF-CALIBRATED DENOISER

Area of Science:

• Medical Imaging
• Signal Processing
• Machine Learning

Background:

• Plug-and-play (PnP) methods utilize application-specific denoisers for inverse problems like MRI reconstruction.
• Training these denoisers often requires substantial data, which is frequently unavailable for specific applications.

Purpose of the Study:

• To propose a novel PnP-inspired method for image reconstruction that bypasses the need for external training data.
• To develop a self-supervised approach for training the denoiser using only the acquired incomplete measurements.

Main Methods:

• The proposed method, Recovery with a Self-Calibrated Denoiser (ReSiDe), trains the denoiser iteratively from image patches within the reconstruction process.
• Both denoiser training and denoising subroutine calls are integrated into each iteration of the PnP algorithm.
• Validation involved comparing ReSiDe against compressed sensing and a PnP method with BM3D denoising on single-coil MRI brain data.

Main Results:

• ReSiDe demonstrates a progressive refinement of the reconstructed MRI images through its iterative self-supervised training approach.
• The method successfully reconstructs images without requiring a separate dataset for denoiser training.
• Comparative analysis showed the performance of ReSiDe against established reconstruction techniques.

Conclusions:

• ReSiDe offers a viable solution for MRI reconstruction when training data is scarce, leveraging a self-supervised denoiser.
• The method's PnP-inspired framework enables efficient and progressive image refinement.
• This self-supervised approach advances the applicability of PnP methods in data-limited inverse problems.