Exploring GAN-Based Synthetic Medical Imaging for Improved Tumour Detection and Diagnosis

Abstract

Medical imaging plays a critical role in tumour detection and diagnosis, but the scarcity of annotated datasets limits the effectiveness of deep learning models. This research investigates the use of Generative Adversarial Networks (GANs) for synthetic medical image generation to improve tumour detection performance. Four GAN models—Vanilla GAN, DCGAN, Conditional GAN (cGAN), and CycleGAN—were implemented to generate high-quality synthetic MRI images and augment existing datasets. Experimental results demonstrate that the integration of synthetic images significantly enhances tumour classification. The cGAN model achieved the highest detection performance with 87% accuracy, F1-score 0.86, and improved representation of rare tumour types. DCGAN produced high resolution images with defined lines of the tumours with 85% accuracy and SSIM 0.88, whilst CycleGAN produced image conversion across modalities with 84% accuracy and PSNR 28 dB. V orig GAN had moderate improvements of 79% accuracy and FID 35.8 which indicates its weak albeit sound foundation. These results underscore that synthetic imaging with GAN has a capability to deal with data set difference, its generalization aspect, and strengthening of tumour detection systems. The research has identified possible opportunities in the advanced generative models in the clinical field, and the need to consider more multimodal synthesis and the optimization of architectures in GANs may lead to their further use in healthcare..