Domain Progressive Low-dose CT Imaging Using Iterative Partial Diffusion Model
Feiyang Liao,Yufei Tang,Qiang Du,Jiping Wang,Ming Li,Jian Zheng
DOI: https://doi.org/10.1109/tmi.2024.3492260
IF: 10.6
2024-01-01
IEEE Transactions on Medical Imaging
Abstract:Traditional deep learning reconstruction (DLR) methods have been sparsely applied in practical low-dose computed tomography (LDCT) imaging, as they heavily rely on the similarity between the latent distributions of data features. However, in real LDCT imaging scenarios, the distribution of data features is highly diverse and complex, which limits the generalizability of existing DLR methods. Recently, diffusion models have shown great potential in the field of LDCT imaging, and some early studies have used them to address the domain generalization problem. However, they still face challenges such as high time consumption, difficulties in training with high resolution, and performance degradation in denoising scenario. In this paper, we propose a novel domain progressive LDCT imaging framework with an iterative partial diffusion model (IPDM) as the core. Firstly, the derived IPDM theoretical framework supports completing the denoising task by iterating a small part of the complete diffusion model, utilizing the strong generation ability of the diffusion model while alleviating time consumption and convergence difficulties. Secondly, a derived condition guided sampling method alleviates sampling bias caused by deviations of the predictive data gradient and Langevin dynamics. Finally, an adaptive weight strategy based on pixel-wise noise estimation can gradually adjust guided intensity. Extensive testing on diverse datasets reveals that our method outperforms traditional iterative reconstructions, unsupervised, and some supervised DLR methods in visual and quantitative evaluations, closely matching the performance of state-of-the-art supervised DLR techniques. Additionally, our IPDM was trained using practical normal-dose CT data, rather than the tested LDCT data. This enables our method to have better generalization ability compared to traditional DLR methods in practical imaging scenarios. Source code is available at https://github.com/LFY1998/IPDM-PyTorch.