PTDM: modelo de difusión guiado por texto para la reconstrucción de TC en dosis bajas

Low-dose computed tomography (LDCT) is critical for reducing radiation risks, but it inevitably introduces noise and artifacts, impairing diagnostic accuracy. While diffusion models have shown potential in LDCT reconstruction, existing methods often face challenges such as insufficient simulation of complex physical noise processes, leading to suboptimal performance and poor generalization. To address these issues, we propose a text-guided Phase Transition Diffusion Model (PTDM) to resolve noise-artifact removal and mismatches in physical processes. First, we construct physics-driven phase-transition diffusion mechanism, with the coupling state of LDCT and Gaussian noise as the diffusion endpoint, explicitly modeling the Poisson-Gaussian noise process and optimizing the noise balance point through phase-transition theory. Second, we introduce multi-modal semantic constraints to suppress over-smoothing effects through Contrastive Language- Image Pre-training (CLIP) loss. Finally, we incorporate wavelet frequency-domain loss to enhance high-frequency structural fidelity, further improving image quality. The superiority and robustness of our method are validated across multiple datasets (Mayo 2016/CHAOS 2019/Nut) and dose levels.