ENHANCING BRAIN MRI QUALITY: FROM NOISE REDUCTION TO IMAGE SYNTHESIS USING GENERATIVE ADVERSARIAL NETWORKS

Convolutional Neural Networks (CNNs) excel in computer-assisted diagnosis, given sufficient annotated data. However, scarcity and fragmentation of medical imaging datasets limit model performance. Generative Adversarial Networks (GANs) address this challenge by creating realistic additional training images to complement existing datasets. Although previous studies explored noise-to-image or image-to-image GANs separately, the potential synergy of combining both approaches remains largely unexplored. Here, we propose a novel two-step GAN-based data augmentation (DA) method for enhancing brain MRI datasets, covering both tumor-inclusive and tumor-exclusive scenarios. Firstly, Progressive Growing of GANs (PGGANs) generates high-resolution MRI images with diverse characteristics. Subsequently, Multimodal UNsupervised Image-to-image Translation (MUNIT) further refines texture and shape of PGGAN-generated images, aligning the images closely with real MRI scans. Our approach significantly improves CNN-based diagnosis across various medical imaging tasks, demonstrating its effectiveness with limited and fragmented datasets.