Analyzing and Advancing Unpaired PET to CT Image Translation: Improvements through Enhanced CycleGAN and CUT

Abstract

Medical imaging plays a crucial role in healthcare, with PET and CT scans providing complementary information about metabolic activity and detailed anatomical structures, respectively. However, the reliance on paired PET-CT data poses challenges in terms of data availability, cost, and clinical variability. This research addresses the unpaired translation of PET images to CT images. Motivated by the potential for improved decision-making and the cost-effective utilization of unpaired data, this work aims to overcome the limitations of current medical imaging approaches & to address the effectiveness of this task. The study explores the use of unpaired image to image translation models for the given task including - CycleGAN, CUT, ,UNIT, MUNIT, SCAM-GAN etc & proposed improvements with the current state of the art result for this specific task. The literature review highlights the effectiveness of these models in various domains, emphasizing their loss functions, including adversarial, identity, cycle consistency & perceptual losses, adding attention mechanism and so on. To enhance the model’s capabilities, the research incorporates attention mechanisms aims to capture dependencies in images. The study evaluates the proposed methodology’s accuracy and applicability by comparing results with existing models. The research team aims to demonstrate the feasibility of solving this task by evaluating different state of the art unpaired models & proposes incorporations which gives better accuracy than the current models.