Frequency-Adaptive Adversarial Diffusion for Unpaired CT-to-CTA Image Synthesis

Abstract

Cross-modality translation between computed tomography (CT) and computed tomography angiography (CTA) images is of great significance for reducing radiation exposure and economic costs while avoiding contraindications associated with contrast agents. In this work, we propose a Frequency-Adaptive Adversarial Diffusion Model (FADM) for unpaired CT-to-CTA image translation. Specifically, FADM employs a cycle-consistent non-diffusion module to generate pseudo-CT images that provide conditional guidance for the diffusion model. An adversarial projection mechanism is also introduced to enable fast and accurate sampling with large diffusion step sizes. Furthermore, an adaptive wavelet fusion mechanism is designed to integrate information from the source and target domains according to their similarity in different frequency bands. This mechanism preserves anatomical structures while enhancing vascular details and transforming intensity characteristics toward the CTA domain. Experimental results on clinical lung CT and CTA images demonstrate that FADM achieves superior performance compared with representative GAN and diffusion methods, obtaining a PSNR of 30.34 dB, an SSIM of 0.793, and an FID of 86.81. These results indicate that FADM can effectively preserve anatomical structures and enhance vascular details, providing a promising solution for high-fidelity CT-to-CTA image synthesis.