Abstract
Accelerated Cardiovascular Magnetic Resonance (CMR) image reconstructionremains a critical challenge due to the trade-off between scan time and imagequality, particularly when generalizing across diverse acquisition settings. Wepropose GENRE-CMR, a generative adversarial network (GAN)-based architectureemploying a residual deep unrolled reconstruction framework to enhancereconstruction fidelity and generalization. The architecture unrolls iterativeoptimization into a cascade of convolutional subnetworks, enriched withresidual connections to enable progressive feature propagation from shallow todeeper stages. To further improve performance, we integrate two loss functions:(1) an Edge-Aware Region (EAR) loss, which guides the network to focus onstructurally informative regions and helps prevent common reconstructionblurriness; and (2) a Statistical Distribution Alignment (SDA) loss, whichregularizes the feature space across diverse data distributions via a symmetricKL divergence formulation. Extensive experiments confirm that GENRE-CMRsurpasses state-of-the-art methods on training and unseen data, achieving0.9552 SSIM and 38.90 dB PSNR on unseen distributions across variousacceleration factors and sampling trajectories. Ablation studies confirm thecontribution of each proposed component to reconstruction quality andgeneralization. Our framework presents a unified and robust solution forhigh-quality CMR reconstruction, paving the way for clinically adaptabledeployment across heterogeneous acquisition protocols.