Unsupervised Medical Image Translation with CycleGAN: Applications and Clinical Boundaries

School of Software Engineering, Shandong University, Jinan, China

^* Corresponding author: This email address is being protected from spambots. You need JavaScript enabled to view it.

Abstract

CycleGAN effectively mitigates the critical challenge of paired data scarcity in medical image-to-image translation. This comprehensive review examines its prominent applications across diverse modalities, particularly highlighting its success in MRI-to-CT synthesis for radiotherapy planning (e.g., reducing mean absolute error from 69.29 HU to 29.85 HU), ultrasound speckle noise reduction, and X-ray denoising. The review provides a comparative analysis with established methods, such as Pix2Pix (e.g., SSIM 0.82 vs. 0.85) and DualGAN, to underscore its strengths in robustness when leveraging unpaired datasets. Furthermore, the review identifies significant barriers to clinical deployment, primarily encompassing potential risks to diagnostic reliability induced by generative artifacts—such as creating clinically misleading features or missing subtle pathologies—and the unresolved ethical boundaries concerning the use of synthetic data. Future research directions should prioritize the development of robust artifact reduction techniques, the execution of large-scale multi- center clinical validation studies, and the advancement of hybrid models integrating anatomical priors.