Super-Resolution based Coronary Magnetic Resonance Angiography Acceleration with Enhanced Perceptual Quality

Coronary magnetic resonance angiography (CMRA) is a promising technique for diagnosis of coronary artery disease. The long acquisition time of high-resolution CMRA limits its clinical application. While image super-resolution techniques can reduce acquisition times by reconstructing the spatial resolution of undersampled CMRA data, existing methods pay insufficient attention to the perceptual quality of reconstructed CMRA. To address this, we propose a Coronary Super-Resolution Generative Adversarial Network (CSRGAN), designed to accelerate CMRA acquisition while ensuring high perceptual quality. We introduce the Residual Dense Channel Attention Block (RDCAB) as the basic module of the generator, a relativistic discriminator and a more efficient perceptual loss in CSRGAN. The proposed CSRGAN was compared with the state-of-the-art Super-Resolution Generative Adversarial Network (SRGAN) on retrospectively undersampled data. The results demonstrates that the proposed CSRGAN achieved better image perceptual quality with more realistic and natural details compared with SRGAN in both super-resolution CMRA images and coronary curved planar reformation. Moreover, CSRGAN effectively reconstructs the coronary artery morphology, achieving fine details comparable to that of fully sampled data.