K-space Guided Under-sampling Pattern Optimization for Implicit Neural Representation-based Reconstruction

Motivation: Given a specific acceleration rate, there exist variable under-sampling patterns, which may lead to significant differences in reconstruction performance. It is unclear which under-sampling pattern is the most efficient for the given anatomy and reconstruction method employed. Goal(s): This study aims to develop a data-driven and model-specific optimization approach to optimize under-sampling patterns, thereby improving MRI reconstruction quality. Approach: We proposed an approach based on the stochastic greedy algorithm incorporated with k-space guidance for under-sampling pattern optimization. Results: Our method demonstrated superior reconstruction quality, producing fewer artifacts compared to conventional patterns. The optimized patterns are data-dependent, varying with image anatomy and coil settings. Impact: This study significantly improves MRI reconstruction quality by optimizing under-sampling patterns, potentially leading to more accurate clinical diagnoses. It opens avenues for further research on the adaptability of the under-sampling patterns in various clinical contexts, ultimately enhancing patient outcomes.