Dose-aware Diffusion Model for 3D Low-count Cardiac SPECT Image Denoising with Projection-domain Consistency

作者：Huidong Xie, Wenxia Gan, Xiongchao Chen, Bo Zhou, Qian Liu, Ming Xia, Xiaoyun Guo, Y.-H. Liu, Hongyu An, Ulugbek S. Kamilov, Ge Wang, Albert J. Sinusas, Chunlei Liu · 年份：2024 · DOI：10.1109/nss/mic/rtsd57108.2024.10655170 · 被引用次数：2 · 研究领域：Medical Imaging Techniques and Applications、Advanced MRI Techniques and Applications、Radiomics and Machine Learning in Medical Imaging

SPECT imaging has been widely used in cardiology studies. Since SPECT scans are accompanied by radiation exposure, reducing the injected dose in SPECT scans is an important topic. Deep learning techniques have been investigated for low-dose SPECT imaging. However, previous neural networks proposed for low-dose SPECT imaging have limited generalizability to other noise levels due to different noise amplitude and variances. Recently, diffusion model demonstrated its potential for superior medical image denoising performance and generalizbility to different noise-levels. However, the stochastic nature of diffusion models results in distorted anatomical structures in challenging cases. Here, to address these limitations, we developed DDSPECT-3D, a dose-aware diffusion model for 3D low-count SPECT image denoising. During model testing, we proposed to incorporate MLEM iterative reconstruction updates within the reverse sampling steps to enforce a projection-domain consistency. Tested on 100 patient studies with low-count levels ranging from 5% to 50%, presented results showed that DDSPECT-3D can consistently produce promising denoised results regardless of input noise-levels.