MR Spatiospectral Reconstruction Integrating Subspace Modeling and Self-Supervised Spatiotemporal Denoising

Area of Science:

• Medical Imaging
• Spectroscopy
• Computational Science

Background:

• Magnetic Resonance Spectroscopic Imaging (MRSI) data is often corrupted by significant noise.
• Existing subspace-based methods face challenges in effectively handling high-dimensional spatiospectral noise.
• Accurate reconstruction is crucial for reliable quantitative analysis in MRSI.

Purpose of the Study:

• To develop an advanced reconstruction method for highly noisy MRSI data.
• To integrate subspace modeling with a learned spatiotemporal denoiser.
• To improve the accuracy and performance of MRSI reconstruction.

Main Methods:

• A novel method integrating subspace modeling and a pre-learned spatiotemporal denoiser was developed.
• A self-supervised learning strategy trained the denoiser to differentiate signals from noise.
• An iterative reconstruction framework using Plug-and-Play (PnP)-ADMM synergized subspace constraints, the denoiser, and the encoding model.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art subspace-based techniques.
• Evaluations using numerical simulations and in vivo data confirmed improved reconstruction quality.
• Theoretical analysis supported the benefits of combining subspace projection and iterative denoising.

Conclusions:

• The integration of self-supervised denoising priors and low-dimensional representations offers significant potential for high-dimensional imaging.
• This approach effectively addresses noise challenges in MRSI reconstruction.
• The developed method provides a robust framework for enhanced MRSI data analysis.