HARA-GAN: Hybrid Attention and Relative Average Discriminator Based Generative Adversarial Network for MR Image Reconstruction

Reconstructing high quality and consistent Magnetic Resonance images (MRI) from low under sampled k- space data were a common challenges in medical image construction and analysis. To overcome these challenges, it needs an advanced reconstruction technique. Currently, various deep learning techniques are outperforming traditional medical image reconstruction techniques, like parallel imaging and compressed Sensing techniques. However these techniques performed well in high under sampled data, they will produce low performance when dealing with low under sampled k-space data. In this work, we proposed a generative adversarial network (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. The zero filling image was generated using 12.5% under sampling rates for cartesian, and 10% under sampling rates for both radial and spiral under sampling masks. In generator module the proposed method used two residual U-net with hybrid attention mechanism. In encoder parts of the U-net, we employed an spatial and self attention and in the decoder side we have employed spatial and channel wise attention techniques. The reconstructed MR image quality and consistency was evaluated using peak signal-to-noise ratio (PSNR), structural similarity index measures (SSIM), and normalized mean square error (NMSE). The results indicated that HARA-GAN outperforms DAGAN, RefineGAN and RSCA-GAN methods based on error maps and quantitative evaluation metrics in terms of both image quality and consistencies on two prominent data sets, brain and knee MR image reconstruction. Over all, HARA-GAN have important implications for clinical applications, as high-quality and consistent results are crucial for accurate diagnosis and treatments of brain and knee MR images.