Learnable Objective Image Function for Accelerated MRI Reconstruction

Magnetic Resonance Imaging (MRI) provides strong contrast for soft tissues but requires long acquisition times, oftentimes resulting in the motion artifacts. Recent advancements in MRI reconstruction from undersampled data rely on compressed sensing (CS) and deep learning (DL) techniques, allowing for significant scan acceleration while maintaining the image quality nearly at the fully sampled level. In this study, we propose to use the convolutional neural networks (CNN), such as U-net, to parametrize the objective function employed in the compressed sensing optimization problems. By doing so, our aim is to avoid unrealistic reconstructions often associated with traditional DL-based image reconstruction techniques. To validate the proposed method, we used the CMRxRecon dataset containing cardiac raw k-space data. The results demonstrate realistic reconstruction of anatomical structures, outperforming classical CS reconstruction methods.