K-space replacement layer for improving image sharpness in machine learning reconstruction

Motivation: Machine learning reconstruction (MLR) is a state-of-the-art method for reconstructing MR images from undersampled data. MLR images may exhibit blur, especially when the acceleration is pushed too high. Goal(s): The goal of this study is to improve the image sharpness while maintaining high acceleration. Approach: In this study, we report a new architecture for MLR by incorporating a k-space replacement layer in the network training. Results: We demonstrate that the proposed method can improve image sharpness at an acceleration factor of 4x. Impact: The proposed method has the potential to increase clinical throughput by reducing scan time while maintaining high image quality.