Multi-GAN Aggregation with Style Enhancement for Improved Synthetic Brain Tumor Image Generation

Building upon recent advancements in deep learning for medical image analysis, particularly generative adversarial networks (GANs), we propose a novel approach for synthetic brain tumor image generation. Our work addresses the limitations of single GAN models, which may capture localized features but struggle with broader image context. We introduce Multi-GAN Aggregation with Style Enhancement (MGASE), which leverages the strengths of diverse GAN architectures: DCGAN, WGAN, and StyleGAN2. Inspired by the AGGrGAN model, MGASE aggregates the outputs from these GANs. This aggregation may also enable us to capture the distributed image features, leading to more accurate and diverse synthetic tumor representations. Moreover, we have used style transfer to refine the aggregated image and ensure faithful resemblance to real brain scans. We evaluated MGASE on the BraTS 2020 dataset, and according to our experiments, MGASE generated images with an SSIM value of 0.89 and a PSNR value of 28.33, which we feel could be improved further since we had some resource constraints. This demonstrates its superiority over existing methods in terms of structural similarity and perceptual quality. Our results highlight the effectiveness of combining diverse GANs and style transfer for generating highly realistic and valuable synthetic brain tumor images, catering to the common issue of data unavailability in the healthcare domain. This opens the door further for improved medical diagnosis and treatment planning.