Self-supervised Deep Unrolled Model with Implicit Neural Representation Regularization for Accelerating MRI Reconstruction

TL;DR

This work tackles the challenge of accelerating MRI by eliminating the need for large fully-sampled training data. It introduces UnrollINR, a zero-shot self-supervised unrolled reconstruction that integrates an implicit neural representation as an explicit regularizer within a physics-guided framework, with data fidelity enforced by a CG-based DC step. The approach achieves superior reconstruction quality at high acceleration rates compared with supervised and self-supervised baselines, while reducing training cost through a compact architecture. The method demonstrates robustness across datasets and undersampling patterns, suggesting strong practical potential for fast, scan-specific MRI reconstruction.

Abstract

Magnetic resonance imaging (MRI) is a vital clinical diagnostic tool, yet its application is limited by prolonged scan times. Accelerating MRI reconstruction addresses this issue by reconstructing high-fidelity MR images from undersampled k-space measurements. In recent years, deep learning-based methods have demonstrated remarkable progress. However, most methods rely on supervised learning, which requires large amounts of fully-sampled training data that are difficult to obtain. This paper proposes a novel zero-shot self-supervised reconstruction method named UnrollINR, which enables scan-specific MRI reconstruction without external training data. UnrollINR adopts a physics-guided unrolled reconstruction architecture and introduces implicit neural representation (INR) as a regularization prior to effectively constrain the solution space. This method overcomes the local bias limitation of CNNs in traditional deep unrolled methods and avoids the instability associated with relying solely on INR's implicit regularization in highly ill-posed scenarios. Consequently, UnrollINR significantly improves MRI reconstruction performance under high acceleration rates. Experimental results show that even at a high acceleration rate of 10, UnrollINR achieves superior reconstruction performance compared to supervised and self-supervised learning methods, validating its effectiveness and superiority.

Figures (7)

• Figure 1: Overview of the proposed UnrollINR framework with an unrolled iterative architecture incorporating an INR-based network as the regularization term.
• Figure 2: Comparative results of different methods on the fastMRI knee dataset under various acceleration rates R. The quantitative metrics PSNR and SSIM are indicated at the bottom of each reconstructed image. Local magnified views of the reconstructed images and absolute error maps are provided. The undersampling masks used for the corresponding acceleration rates R are displayed on the far left of the figure.
• Figure 3: Comparative results of different methods on the fastMRI brain dataset under various acceleration rates R. The quantitative metrics PSNR and SSIM are indicated at the bottom of each reconstructed image. Local magnified views of the reconstructed images and absolute error maps are provided. The undersampling masks used for the corresponding acceleration rates R are displayed on the far left of the figure.
• Figure 4: Comparative results of prospective reconstruction for all methods at an acceleration rate R of 8. Locally magnified views of each reconstructed image are provided. The undersampling mask used in the prospective experiment, with an ACS size of 26, is displayed on the leftmost side of the figure.
• Figure 5: The impact of two key hyperparameters in UnrollINR on PSNR values using the fastMRI knee dataset with an acceleration rate of R = 10.