Advancing Self-Supervised Learning for Highly Accelerated MRI Reconstruction Through Parallel Imaging Consistency

Motivation: To improve training for self-supervised deep learning (DL) reconstruction in highly accelerated acquisition scenarios. Goal(s): We present a self-supervised approach that assesses the quality in both k-space and image domain, drawing on consistency ideas from parallel imaging. Approach: Parallel imaging consistency is achieved through carefully crafted perturbations for R-fold acceleration, designed to be restorable with parallel imaging reconstruction. Outputs for both perturbed and unperturbed inputs are analyzed and used in conjunction with k-space masking. Results: Proposed method achieves significant aliasing reduction at R=6 and R=8, outperforming state-of-the-art self-supervised methods on fastMRI dataset. Impact: This work proposes an improved training strategy for self-supervised MRI reconstruction by applying well-designed perturbations to input images. This ensures alignment with parallel imaging techniques and reduces aliasing artifacts, achieving visible improvements at high acceleration rates.