Multi-phase liver CT virtual contrast enhancement and diagnosis based on phase-aware diffusion network

Multi-phase contrast-enhanced CT (CECT) plays a crucial role in liver tumor diagnosis but is limited by additional radiation exposure and a restricted patient population. While virtual contrast enhancement (VCE) methods can generate virtual liver CECT images, challenges persist due to detail-deficient non-contrast CT (NCCT) images and phase-wise feature variations. In this paper, we propose phase-aware diffusion (PA-Diff), a novel model for multi-phase liver CT enhancement and contrast-agent-free liver tumor diagnosis, which combines two novel designs: 1) Phase-specific path allocation that enables precise multi-phase enhancement by combining independent diffusion paths and shared encoder and decoder; 2) Cycle re-encoding learning that enhances tumor detail by recovering target images and re-encoding them into complementary features for loss calculation during training. We also perform automatic diagnosis on the VCE CT images. Experiments on two dataset demonstrate effective multi-phase enhancement and accurate Hepatocellular Carcinoma (HCC) screening and Microvascular Invasion (MVI) diagnosis, highlighting the model's clinical value.