Hybrid Attention and Relative Average Discriminator Based Generative Adversarial Network for MR Image Reconstruction

Abstract

Reconstructing high-quality and consistent Magnetic Resonance Images form low under sampled k- space data were a common challenging in medical image analysis. It needs an advanced reconstruction technique to overcome those challenges. Currently, deep learning techniques outperform traditional methods such as parallel imaging and Compressed Sensing for MR image reconstruction. The DAGAN, RefineGAN, RCA-GAN and RSCA-GAN worked well in high under-sampled data, and rarely show reasonable performance when dealing with low under-sampled k-space data. In this study, we proposed a GAN-based architecture called HARA-GAN by integrating residual U-net with Hybrid attention and Relative average discriminator, to modify MRI reconstruction and reduce the noise caused by low under-sampling rates. In the generator module, the proposed method used two residual U-net with a hybrid attention mechanism. In the encoder parts of the residual U net, it used spatial and self-attention, and in decoder side also used spatial and channel wise attention. The proposed HARA-GAN was compared to the zero filling and RSCA-GAN on brain and knee MRI datasets using both cartesian and non-cartesian under-sampling masks. It used 12.5% sampling rates for cartesian, and 10% sampling rates for both radial and spiral under-sampling masks. The reconstructed image quality and consistency were evaluated using the peak signal-to-noise ratio (PSNR), structural similarity index measures (SSIM), and normalized mean square error (NMSE). The results indicated that HARA-GAN outperforms states of arts MR image reconstruction methods based on error maps and quantitative evaluation metrics in terms of both image quality and consistency. It improved the PSNR values for brain from 30.49 to 32.60, and for knee 32.08 to 35.89 for cartesian low under sampling rates. Thus, HARA-GAN has important implications for clinical applications, as high-quality and consistent results are crucial for accurate diagnosis and treatments of brain and knee MR images.