Unpaired Image-to-Image Translation for High-Accuracy CTA Generation from CT Images Using Cycle GAN

Abstract

This study explores the application of CycleGAN, a variant of Generative Adversarial Networks (GANs), for generating Computed Tomography Angiography (CTA) images directly from CT scans. Traditional CTA methods involve risks associated with contrast agents and radiation, prompting the need for non-invasive alternatives. CycleGAN facilitates unpaired image-to-image translation, bypassing the requirements for paired datasets typically used in medical imaging. Using a dataset comprising CT scans from 87 patients and CTA scans from 56 healthy aortas, our study employed advanced preprocessing techniques to optimize the CycleGAN performance. Quantitative evaluations demonstrate the model's efficacy with a Structural Similarity Index (SSIM) of 0.73 ± 0.09 and a Filtered Peak Signal-to-Noise Ratio (Filtered PSNR) of 18.43 ± 0.96, indicating robust image synthesis capabilities. Our Study advances medical imaging by automating CTA image synthesis, potentially improving diagnos-tic accuracy and patient care while mitigating procedural risks. Future research directions include refining model architectures, expanding datasets, and validating synthesized images across broader clinical contexts.