MRI Super-Resolution using Generative Adversarial Network and Discrete Wavelet Transform

Deep Learning Approaches have brought in major advances in super-resolution. Deep Learning architectures like convolutional neural networks or auto-encoders are trained to beget high-resolution (HR) images from the low-resolution (LR) images. Discrete wavelet transformations (DWT) is used to bring out the high frequency (HF) sub-bands of the image, which is further used to reduce the error while synthesizing super-resolution images. The proposed architecture combines the features generated by the DWT and Super-Resolution Generative Adversarial Networks (SRGAN) to produce images of high resolution. The implemented topology initially applies DWT to the image and separates out the image into two fundamental bands (HF and LF sub-bands). The features are then generated by separate generators and then Combined by inverse 2d discrete wavelet transformation (IDWT) which is then passed on to the discriminator to evaluate the generated image. A key feature of this design is that each feature is given importance and is generated by separate generators which makes the reconstruction and improvising the quality of the image more feasible. From the experiments conducted on medical images such as magnetic resonance imaging (MRI), the proposed design is computationally simpler and yet produces competitive and often improved results than state-of-the-art alternatives in terms of peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM).