Enhancing brain tumor MRI scans using GAN-based image synthesis and deep CNN classification

Magnetic Resonance Imaging commonly known as MRI, has been pivotal since ages when it comes to diagnosis of disorders related to neurological issues. However, their noise, low resolution and artifacts has compromised its effectiveness. Traditional image enhancement techniques struggle to maintain fine details in MRI scans, leading to suboptimal diagnostic accuracy. Generative Adversarial Networks (GANs) have now become an apparent high-powered and an influential deep learning approach for enhancing MRI images by reducing noise and a convincing and improved structural fidelity. In this study, different GAN variants namely CycleGAN, Pix2Pix, CGAN, ESRGAN and VanillaGAN were evaluated and their performance on the BraTS2021 which is a dataset consisting of has multi-institutional, multi-parametric MRI (mpMRI) scans which are clinically-acquired. The model's performance was assessed considering Mean Squared Error (MSE), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index (SSIM). This research aims on assessing these metrics, CycleGAN and Pix2Pix demonstrate the best performance, achieving lowest MSE values 0.000785 and 0.00065 respectively and highest PSNR and SSIM scores which is 37.07dB and 0.9816 for CycleGAN and 37.89 dB and 0.9785 for Pix2Pix GAN. This study contributes to GAN-based enhancement improves clarity and guiding to reliable visualization and aiding in finer diagnostics. Further, brain tumor dataset from Kaggle was used for the classification of the tumour using Convolutional Neural Networks (CNN’s) and their results were assessed using various evaluation parameters.