Learn Stable MRI Under-Sampling Pattern with Decoupled Sampling Preference

Accelerating the scanning time of magnetic resonance imaging (MRI) and improving the imaging quality is critical to the patient experience in clinical applications. In accelerated MRI, MR data can be under-sampled in the raw k-space, and the available measurements are used for image reconstruction. Many recent studies have shown that the quality of image reconstruction depends largely on the MRI under-sampling pattern. Our study shows that the way to binarize sampling masks in previous deep-learning-based under-sampling patterns leads to unstable reconstruction results when cooperating with data-driven reconstruction methods. Therefore, we design a stable under-sampling pattern learning strategy, which straightly optimizes the sampling mask at fixed k-space locations. In addition, our method designs the sampling pattern from both reconstruction contribution and reconstruction capacity perspectives, which are not considered in previous approaches. We decouple the under-sampling preferences and form a new paradigm for pattern optimization. To accelerate the joint learning process, we utilize prior knowledge of the MRI k-space energy distribution to guide the optimization of under-sampling patterns in our proposed paradigm. Experimental results demonstrate that our proposed under-sampling pattern learning strategy can yield better reconstruction quality than previous methods and facilitate the training process, irrespective of the task being single-coil MRI reconstruction or parallel MRI reconstruction.