A Context-Encoders-Based Generative Adversarial Networks for Cine Magnetic Resonance Imaging Reconstruction

Cine imaging serves as a vital approach for non-invasive assessment of cardiac functional parameters. The imaging process of Cine cardiac MRI is inherently slow, necessitating the acquisition of data at multiple time points within each cardiac cycle to ensure adequate temporal resolution and motion information. Over prolonged data acquisition and during motion, Cine images can exhibit image degradation, leading to the occurrence of artifacts. Conventional image reconstruction methods often require expert knowledge for feature selection, which may result in information loss and suboptimal outcomes. In this paper, we employ a data-driven deep learning approach to address this issue. This approach utilizes supervised learning to compare data with different acceleration factors to full-sampled spatial domain data, training a context-aware network to reconstruct images with artifacts. In our model training strategy, we employ an adversarial approach to make the reconstructed images closer to ground truth. We incorporate loss functions based on adversarial principles and introduce image quality assessment as a constraint. Our context-aware model efficiently accomplishes artifact removal and image reconstruction tasks.