Precision Imaging Of Low-Dose Cerebral CT Perfusion Via Temporal-Spatial Regularization And Deep Learning

Stroke is a brain tissue damage caused by brain ischemia,which is common in middle-aged and elderly people,and has a high recurrence,disability and mortality rate.Cerebral CT perfusion can reflect the blood flow of brain tissue,and has become an indispensable imaging modality in the clinical diagnosis and treatment of stroke.However,the high radiation dose of cerebral CT perfusion exposes patients to a high risk of carcinogenesis.Currently,the most straightforward way to reduce radiation dose is to reduce the tube current(mA)in the scanning protocol.However,the reduction of the tube current causes the detected signal to be interfered by electronic noise and a large number of noise-induced artifacts in the reconstructed image,which will seriously affect the doctor’s diagnosis and treatment of patients.In view of the problem of impaired CT perfusion imaging quality in low-dose cases,two low-dose image restoration methods are proposed:in this paper,as follows:(1)Under the framework of CT statistical iterative reconstruction,this paper proposes a temporal-spatial regular term,which is referred to as "PWLS-ICTGV" for short.Cerebral CT perfusion sequence images can be divided into dynamic part(enhanced information)and static(anatomical information)part.The two parts have a lot of redundant information in both temporal and spatial dimensions.Based on the above knowledge,this paper proposes the ICTGV regular term.The regular term consists of two second-order Tensor Total Generalized Variation(TTGV)equations,and the temporal-spatial constraint strengths of the two are different.The CT perfusion sequence image can be automatically divided into two parts in the process of iterative solution to achieve an optimal balance of local temporal and spatial constraints:the part of higher temporal dimension constraints contains more static(anatomical)information,while the part of the lower time dimension constraint contains more dynamic(enhanced)information.In this paper,the iterative reconstruction framework is optimized by the alternating direction multiplier method,and the influence of different parameter changes on the results in the PWLS-ICTGV method is studied.In addition,in order to verify and evaluate the performance of the PWLS-ICTGV method,both the digital brain phantom and clinical patient data are used in this paper.The qualitative and quantitative results showed that the method can effectively suppress the noise and obtain a more accurate perfusion parameter map.(2)A deconvolution model based on deep learning is proposed,which is referred to as "CTPNet".We uses the Alternating Direction Multiplier Method(ADMM)to optimize the objective function of deconvolution,and expands the substeps of ADMM to construct a deconvolution network.After that,a parametric plug-and-play priori is constructed by using the residual network,to replace the regular term in the original objective function.Finally,in order to fully extract the rich information of normal-dose images,the Arterial Input Function(AIF)and the Time-Density Curve(TDC)at the low-dose case were used as input data,and the perfusion parameter map at the normal-dose case was used as the tag data to train the parameter of the entire deconvolution network.The network structure of the CTPNet method is analyzed and discussed.At the same time,based on the clinical patient data,the performance of the CTPNet method is evaluated.Qualitative and quantitative results indicate that the CTPNet method can obtain high quality perfusion parameter maps from low-dose images.