Research On Artifact Removal Method Of CBCT Image

With the advantages of low radiation dose,high X-ray utilization,unsophisticated Equipment,fast scanning speed,the cone bone computed tomography(CBCT)is more and more widely used.CBCT has been used in the examination of dental bed,chest and pelvis,as well as the detection of precision instruments and equipment.However,compared with the CT system,CBCT system has a lager cone angle in the flat-panel detector and lacks the collimator used to isolate scattering artifacts.Therefore,the CBCT images are seriously affected by scattering artifacts,which limits the application of CBCT in clinical medicine.This thesis aims to explore an effective method to suppress CBCT artifacts.First of all,the thesis explains the background of the topic of CBCT,introduces the common artifact correction methods from three aspects of hardware,software,and hardware and software.Besides,this thesis introduces the CBCT system,including the imaging principle of CBCT system and the description of each component in the system,discuses the causes of artifacts,analyses common image reconstruction algorithms,including iterative and analytical methods.Deep learning provides efficient solutions to the imaging processing,so the related knowledge of deep learning is also introduced.Secondly,the CBCT image without artifact is difficult to obtain,so this thesis makes the data set of chest region based on the CBCT image with artifacts and the CT image without artifact,but this will lead to the problem of data set misalignment.This thesis proposes to introduce contextual loss(CX loss)to solve the problem.Through optimization,we find the optimal network parameters and the weight ratio of the loss function.Finally,the correction results of this method and perceptual loss are evaluated by four indicators: structural similarity,peak signal-to-noise ratio,average absolute error and standard deviation.The results show that contextual loss is more suitable to solve the problem of misalignment of datasets in this thesis and this method can partly suppress artifacts,but the details of the original image are lost,and some areas are blurred.Finally,in order to solve the problem of image detail missing,this thesis aims to find a convolutional neural network that can make full use of image feature information and avoid image detail information losing.In this thesis,an improved residual network is constructed based on the residual network.By adding connection between the residual block and the residual block to make local feature fusion,and pass the fusion feature backwards,so the intermediate feature information from different residual blocks can be fully utilized.Through quantitative analysis,it is found that this method is improved compared with the baseline method.In addition,by using the average CT value,average structural similarity,average peak signal-to-noise ratio,and average absolute error as evaluating indicators to compare the first method and other artifact correction methods,it is found that the results show that this method can better remove the artifacts and retain the internal contour and texture information of the original image.In addition,the test results of head data show that this method also has certain universality in other body parts.