Cross-modal enhanced sparse CT imaging via null-space denoising diffusion with random medical measurement embedding

Abstract

Recent advancements in diffusion models for sparse-view medical computed tomography (CT) have mitigated common issues in supervised deep learning, such as over-smoothing and limited generalization. However, these models often rely on lengthy sampling chains, leading to impractical computation times and error accumulation, especially under significant data distribution shifts. Moreover, they typically overlook clinical noise, which is prevalent in real-world scenarios. To address these challenges, we introduce the Denoising Diffusion model with cross-Modal prior and physical Measurement embedding (DDMM-CT) for reconstructing sparse-view CT images. DDMM-CT refines the null space of intermediate results during inference by leveraging cross-modal geometric information, narrowing the target region in each denoising step. The measurement-related space component is replaced with a combination of the physical operator and measurements to enforce data consistency with minimal additional computation. An error-feedback correction block is integrated to reduce errors from imperfect reconstruction steps. We also present DDMM-CT-noise, designed for clinical scenarios with complex noise mixtures. The proposed method demonstrates superior generalization and flexibility, allowing adjustments in the number of projections and measurement noise intensity without retraining. Our results show that DDMM-CT outperforms recent comparable methods in terms of inference time and image quality. The code is available at https://github.com/Lxy98Code/DDMM-CT.