A Simulation & Technical Feasibility Study for Removing Blur Artefacts from Emission Tomography Images Using Generative Adversarial Networks

Nuclear imaging holds significant importance, offering valuable insights into the physiological and functional aspects of organs and tissues, aiding in the diagnosis, treatment planning, and monitoring of various medical conditions. However, there artefacts arise when the subject immobility during data acquisition is violated. Such artefacts can result in the misinterpretation of images, leading to incorrect diagnoses or missed abnormalities. Generative Adversarial Networks (GANs), a type of deep neural network, have demonstrated their ability to effectively model underlying patterns and outperform traditional methods in various tasks, including image deblurring. In this study, we propose an Enhanced Super-Resolution Generative Adversarial Networks (ESRGAN) based method to address the issue of blur in SPECT and PET images. We applied blur effect including rigid motion effect to SPECT and PET images with minimal involuntary motion. The input to the GAN model was blurred SPECT and PET images, downscaled by a factor of four. The objective of the GAN model was to generate the original, unblurred image. Our proposed model achieved a 0.95 SSIM score on the test dataset. The proposed method for motion correction in SPECT and PET images not only addresses the pressing issue of motion artefacts but also holds the potential to improve the accuracy and reliability of diagnostic imaging significantly.