Dynamic Cerebral Perfusion CT Imaging via a Low-Dimensional Manifold Model

Dynamic cerebral perfusion computed tomography (DCPCT) enables the assessment of cerebral blood circulation and is therefore widely applied to cerebrovascular diseases diagnosis. However, the scanning protocol of this imaging technique significantly elevates the radiation exposure dose for patients. While decreasing the x-ray tube current directly reduces the radiation dose, the correspondingly increased quantum noise in projection data will severely degrade the quality of DCPCT images reconstructed using the classical filtered back-projection (FBP) algorithm. To enhance the image quality of DCPCT under low-dose scenarios, this study presents a statistical iterative reconstruction method based on the penalized weighted least-squares (PWLS) criterion, with a low-dimensional manifold model (LDMM) for regularization, called PWLS-LDMM. This approach leverages the observation that DCPCT images exhibit low-dimensional structure in their patch manifolds. By incorporating the dimension of these patch manifolds as regularization constraints into the image reconstruction process, subtle structural features can be effectively preserved in low-dose DCPCT images. We further developed an efficient alternating optimization algorithm to minimize the PWLS-LDMM objective function. In the experiments, a digital brain perfusion phantom and patient data were utilized to perform a comprehensive qualitative and quantitative assessment of the PWLS-LDMM method. The results show that compared to several representative model-based reconstruction methods, the presented PWLS-LDMM method exhibits significant advantages in noise reduction, artifact suppression, and structural detail preservation. Moreover, the DCPCT images reconstructed with this approach yield more precise cerebral hemodynamic parameter maps, providing reliable support for downstream clinical applications.