Decoupled Diffusion Model for Medical Image Translation

Dechao Qiu, Xiaoming Liu, Zilong Yuan, Jinshan Tang, Tianxiang Lei

Abstract: Medical image translation enables the generation of contrast-enhanced CT (CECT) images from non-contrast CT (NCCT) scans, reducing dependence on iodinated contrast agents (ICAs) and minimizing associated health risks. Although diffusion models outperform generative adversarial networks (GANs) in medical image translation, challenges remain in improving sampling speed and restoring anatomical details. To address these issues, we propose a Decoupled Diffusion Model (DDM). Specifically, we first decouple the input image into high-frequency and low-frequency components via discrete wavelet transform (DWT) to enable parallel computing for accelerated sampling. Furthermore, we design a Wavelet UNet (WUNet) to enhance the recovery of anatomical details by leveraging the multi-scale representation capabilities of wavelet transform. Extensive experiments on two clinical datasets, TAP-CT and Coltea-Lung-CT-100W, demonstrate the superior performance of our method, indicating its potential for real-world clinical translation.

External IDs:dblp:conf/smc/QiuLYTL25