Cross-domain distribution adversarial diffusion model for synthesizing contrast-enhanced abdomen CT imaging

Abstract

Synthesizing contrast-enhanced CT imaging (CE-CT imaging) from non-contrast CT imaging (NC-CT) without the need for chemical contrast agents (CAs) injection holds significant clinical value, as CE-CT imaging plays a crucial role in diagnosing liver tumors, especially in identifying and distinguishing benign from malignant liver tumors. However, challenges within CT imaging, such as the low variability in intensity distribution and limited distribution changes, have hindered the effectiveness of existing synthetic methods, including GAN-based methods and diffusion model (DM)-based methods, in synthesizing CE-CT imaging. We propose a novel cross-domain distribution adversarial diffusion model (AdverDM) for CE-CT imaging synthesis, which overcomes the aforementioned challenges and facilitates the synthesis of CE-CT imaging. Our AdverDM incorporates three key innovations: (1) Cross-domain distribution adversarial learning is introduced into DM, enabling the utilization of cross-domain information to learn discriminative feature representations, addressing the limitations of existing DM based methods in capturing conceptually-aware discriminative features and extracting CA-aware feature representations. (2) A content-oriented diffusion model is creatively designed to guide tissue distribution learning, assisting DM in overcoming the challenge of low variability in intensity distribution. (3) A novel structure preservation loss is proposed to maintain the structural information, avoiding the problem of structural destruction faced by DMs. AdverDM is validated using corresponding two-modality CT images (pre-contrast and portal-venous phases), which is a clinically important procedure that benefits liver tumor biopsy. Experimental results (PSNR: 24.78, SSIM: 0.83, MAE: 6.94) demonstrate that our AdverDM successfully synthesizes CE-CT imaging without the need for chemical CAs injection. Moreover, AdverDM’s performance surpasses that of state-of-the-art synthetic methods.