Image Reconstruction Using Generative Adversarial Networks with a Hybrid Loss
MetadataShow full item record
High-quality magnetic resonance images (MRI) provides detailed anatomical information important for clinical application and quantitative image analysis. In many clinical applications, MRI k-space data is undersampled in order to reduce the scanning time and improve the patient experience. In this case, image reconstruction algorithm becomes important to maintain the image quality using only undersampled data. Deep learning methods have demonstrated great potential in MR image reconstruction due to its ability to learn the non-linearity relationship between the under-sampled k-space data and the corresponding desired image. However, most studies have found that high-frequency details are lost in deep learning based reconstruction and the images are unsatisfactory perceptually with overly smooth textures. Among these methods, Generative Adversarial Networks (GANs) are known to reconstruct images that are sharper and more realistic-looking. In this study, we propose a novel hybrid loss function for GAN to reconstruct high perceptual quality MR images. The experiment demonstrates that the hybrid loss function is superior to the individual ones alone for GAN-based reconstruction.