A text-guided Brownian bridge diffusion model for unified multiphase contrast-enhanced CT synthesis

Abstract

Multiphase contrast-enhanced CT (CECT) is important for evaluating focal liver lesions because enhancement patterns reflect lesion vascularity and liver perfusion. However, some patients cannot safely receive iodinated contrast agents, and repeated multiphase scans increase radiation exposure. These limitations motivate the synthesis of CECT from non-contrast CT (NCCT). Existing NCCT-to-CECT methods typically generate each phase using a separate model or require multiphase inputs, limiting unified and phase-controllable synthesis. To address this problem, we propose the Text-Guided Brownian Bridge Diffusion Model (TGBBDM), a text-conditioned image-to-image diffusion framework based on a Brownian-bridge formulation that generates both arterial (ART) and portal-venous (PV) phases within a single model. Phase-specific text prompts are encoded and injected into the denoiser to guide phase-aware generation. On a retrospective single-center multiphase liver CT dataset, TGBBDM shows overall superior performance across whole-image and liver-ROI evaluations. In particular, relative to a competitive Brownian-bridge-based baseline, it improves whole-image PSNR from 24.14/23.38 to 24.49/23.99 in ART/PV and whole-image PCC from 0.9670/0.9680 to 0.9695/0.9731. It also achieves competitive whole-image SSIM of 0.8457/0.8397. These findings suggest that text-guided, phase-controllable NCCT-to-CECT synthesis may provide useful complementary information for liver-lesion assessment, although larger multi-center studies are still needed before broader clinical use.

Ähnliche Arbeiten