Motivation: Existing zero-shot MRI reconstruction methods struggle to impose effective constraints on non-sampled k-space regions, leading to degraded reconstruction quality. Goal(s): To improve reconstruction quality by effectively handling non-sampled k-space regions. Approach: Introduce AKSM (Attention-based K-space Selective Mechanism) that selectively focuses on critical nonsampled k-space regions with indirect constraints. Results: Our experiments on the FastMRI brain dataset 1 demonstrate superior performance across multiple acceleration factors (4×, 8×) and sampling patterns (Gaussian, uniform), consistently outperforming existing zero-shot methods 2 in both reconstruction quality and computational efficiency. Impact: Our AKSM enables robust zero-shot MRI reconstruction by effectively utilizing undersampled k-space data. This breakthrough allows for significant scan time reduction without compromising image quality, potentially transforming clinical practice and making advanced MRI more accessible to patients.
Joo et al. (Tue,) studied this question.