HALO: High-frequency enhanced dose-aware diffusion model for arbitrary low-dose PET reconstruction.

Abstract

Reconstructing standard-dose PET images from low-dose scans offers a promising approach to mitigating associated radiation risks in PET imaging. In clinical practice, the injected dose level is influenced by various factors such as scanner sensitivity, imaging protocols, and specific diagnostic requirements. Additionally, even with the same injected dose, the "effective dose" manifested in the images vary across different individuals due to difference in body composition. Beyond dose variability, preserving high-frequency details in PET images poses a further challenge, as PET inherently lacks detailed structural information. To address these challenges, we propose a High-frequency enhanced dose-Aware framework for LOw-dose PET reconstruction framework, namely HALO, for reconstructing standard-dose PET images from arbitrary low-dose inputs. First, we introduce a Dose Adaptation module to capture the effective dose presented in the images by integrating injected dose information with the estimated noise level. The captured effective dose is then used to control the reconstruction process. To further enhance high-frequency details, we adopt a high-frequency residual generation strategy, where a pre-trained CNN is used to recover low-frequency components and the diffusion model is dedicated to predicting high-frequency residuals. Accordingly, we introduce a Frequency Information Separator (FIS) and also a High-Frequency Modulator (HFM) to enrich high-frequency generation. Extensive experiments on a public multi-dose PET dataset demonstrate that HALO outperforms state-of-the-art approaches both quantitatively and qualitatively, highlighting its potential for real-world clinical applications.