Research On Artifact Removal Method Of CBCT Image Based On Convolutional Neural Network

The emergence of cone beam CT(CBCT)is a major innovation to the existing computer tomography(CT)technology.The CBCT system offers advantages such as small size,low radiation dose,fast scanning speed,high ray utilization,and the same spatial resolution within and between slices.It has broad application prospects in image-guided radiotherapy(IGRT),chest and abdomen examination,oral 3D reconstruction,and industrial inspection.However,compared with the CT system,the flat-panel detector used in the CBCT system has a larger receiving area and lacks a collimator similar to the traditional CT system to isolate the scattering.Therefore,the scattering artifacts in the CBCT image are more serious,which reduces the accuracy of its application in the IGRT system.This article aims to explore an effective CBCT artifact correction method.Firstly,the paper explains the basic principles of CT and CBCT imaging,discusses the causes and types of artifacts,and studies some traditional CBCT reconstruction algorithms,such as iterative and analytical methods.Convolutional neural network(CNN)provides a new and more efficient solution to the problem space,so the paper also introduces the composition of convolutional neural network in detail.Secondly,the paper adopts an algorithm based on mean squared error(MSE)on the problem of artifact correction.This method constructs a simple 17-layer CNN network to learn the mapping relationship from CBCT images to planned CT images.The pelvic data of 20 patients is preprocessed and send to the network for training.The MSE loss is used to optimize the target function.The experiment prove that the method significantly suppresses the artifacts in the CBCT image,but introduces blur in the smoothing correction process,and loses part of the image details.Finally,for the trouble of correcting image smoothing and blurring,the paper introduces the concept of contextual loss(CX Loss)based on CNN.The main idea is to treat an image as a set of various features,and define the similarity between the images through the matching between the features.Unlike the MSE loss that requires the input image and the label image to be aligned at the pixel level,the spatial position of the feature can be ignored when calculating the CX loss,and local deformation is allowed to a certain extent.At the same time,this paper adjusts the dimensions of the above-mentioned straight-through CNN network to be adaptive,and the network dimension changes with the size of the input image,so that the fusion of features at different levels can be fully utilized.Experimental results show that this method not only removes the scattering artifacts,but also retains the internal contour and texture information of the pelvic region.The proposed algorithm is evaluated with indicators such as average absolute error(MAE),peak signal-to-noise ratio(PSNR),and structural similarity(SSIM).The analysis shows that the method does improve the image quality.The test results of chest data show that this method is also applicable to other body parts and has strong generalization.