Brain MRI image bias correction using generative adversarial network

In magnetic resonance imaging (MRI), one of the most important artifacts is something called intensity inhomogeneity. The presence of a bias field in MRI images has the effect of distorting the genuine pixel value and producing misleading variations in pixel intensity. This artifact has a negative impact on the diagnosis made by radiologists and also has a negative impact on the performance of computer-aided diagnostic techniques such as segmentation. The main aim of this paper is to provide a method for modeling realistic visual noise that is based on a generative adversarial network (GAN). The goal of the model is to improve the performance of a deep network denoiser used for bias correction in MRI images. This paper proposes a modified super-resolution GAN (SRGAN) to reduce the inhomogeneity in the brain MRI images. The discriminator of the model is designed to be very complex to judge the generator more effectively. The generator model has less computational complexity to increase the performance of the model while testing. The use of VGG loss as generator has increased the overall performance of the framework. The proposed modified SRGAN has also been trained specially to increase the brightness and contrast of the output images and obtained better PSNR, MSE and SSIM values. The proposed model helped in improving the brightness and contrast of the resultant images.