Abstract : The utilization of Positron Emission Tomography (PET) in disease diagnosis has been increasing in recent years. The quality of images produced by PET scanners plays a significant role in accurate diagnosis. However, these images often contain substantial noise due to photon attenuation and scatter. Therefore, PET images require attenuation correction (AC) and scatter correction (SC) to provide precise metabolic information about the patient's organs. CT-based correction methods expose the patient to significant ionizing radiation. This study focuses on improving the quality of PET scan images by utilizing Generative Adversarial Networks (GAN), a revolutionary approach in modern medical imaging, to reduce errors and patient exposure to radiation. In this study, 92 epilepsy patients with an average weight of 72.15 kg were scanned. Brain imaging was conducted on the patients following the injection of an average activity of 347.13 MBq of FDG radiotracer over a duration of 1200 seconds. These brain images served as the dataset for our designed algorithm, a GAN-based model. Image quality metrics such as SSIM, PSNR, MSE, FID, and LPIPS were measured. Our AutoGANcoder algorithm, a unique combination of a GAN and an advanced autoencoder, demonstrated that it significantly improved PET imaging quality. When compared to other algorithms, the results show that the AutoGANcoder model is a promising choice for improving brain PET image quality.

Keyword : Brain Images, Deep Learning, Generative Adversarial Network, Positron Emission Tomography.