C2 LOG-GAN: Concave convex and local global attention based generative adversarial network for super resolution MRI reconstruction

Abstract

Medical strategies capture high-resolution images that can be used to diagnose and detect a variety of diseases. The obtained images are luxurious to gain and hard to store, the diagnostic procedure consume a substantial amount of time. High-resolution magnetic resonance imaging provides clean structural data, but it is time-consuming. Generative adversarial networks (GAN) have been commonly used for super image resolution to improve the quality of low-resolution images into high-resolution images. Due to the reliance on observational information from other images, the models are less effective in handling the unique complexities, resulting in compromised high-resolution images. To overcome this issue, a novel deep learning based C2 LOG-GAN is proposed for generating LR to SR images using BRATS dataset. Initially, Concave & Convex Sampling is introduced in the first GAN that generates the low- high-resolution image from the Low resolution (LR) image. Then, Local & Global Sampling is implemented in the second GAN to generated patches with saliency map, and detect the Super resolution (SR) image from the high-resolution image. GD-CNN is used to classify the brain abnormalities from the SR images. The performance of the proposed C2 LOG-GAN has been assessed utilizing L1 loss, L2 loss, peak signal-to-noise ratio (PSNR), Mean Squared Error (MSE), and structural similarity index measure (SSIM) metrics. From the experimental analysis, the proposed C2 LOG-GAN model achieves the high PSNR values of 38.95 % and the proposed GD-CNN obtains 99.66 % of accuracy rate in brain disease classification, which is comparatively high than the state-of-the-art methods.