Enhancing the Performance of Generative Adversarial Networks for the Generation and Reconstruction of Brain MRI Images Using Deep Learning

Brain magnetic resonance imaging (MRI) is a critical tool in diagnosing neurological disorders such as tumors, multiple sclerosis, and Alzheimer's disease due to its ability to provide detailed anatomical information. However, challenges such as image noise caused by patient motion, hardware limitations, and the scarcity of diverse, labeled training data compromise the quality and accuracy of image analysis. This study proposes a novel hybrid architecture based on Generative Adversarial Networks (GANs) to enhance the resolution and quality of brain MRI images. The architecture integrates Conditional GAN (CGAN), StyleGAN, a Denoising Autoencoder, and transfer learning with pre-trained models such as VGG۱۹ and ResNet۵۰. Designed to address issues like noise, data scarcity, and the complexity of brain structures, the model leverages MRI data from reputable datasets, including OpenNeuro (۳,۰۰۰ structural and functional scans) and Kaggle (۱,۵۰۰-۵,۰۰۰ images with diverse pathologies). The data underwent preprocessing steps, including intensity normalization, data augmentation, and spatial alignment. The proposed model, utilizing a hybrid generator for targeted image synthesis, a multi-layered discriminator for precise structural analysis, and a denoising autoencoder for noise reduction, outperformed baseline models (simple Autoencoder, Vanilla GAN, and cGAN). Quantitative evaluation yielded SSIM=۰.۸۸, PSNR=۲۹.۶ dB, and FID=۲۸.۴, indicating high structural similarity and effective noise reduction. Qualitative analysis, conducted by radiologists, confirmed accurate and coherent reconstruction of critical brain structures such as gray matter, ventricles, and the corpus callosum. By delivering high-quality images with reduced artifacts, this model demonstrates significant potential for diagnostic and research applications in medical imaging, contributing to improved detection of neurological disorders.

کلیدواژه ها:

Generative Adversarial Networks ، Brain MRI Images ، Deep Learning ، Denoising Autoencoder ، Transfer Learning ، Image Reconstruction