Research On Denoising Algorithm Of Medical Cone Beam CT Image Based On Convolutional Neural Network

There is no doubt that Computed Tomography(CT)is playing an increasingly important role in the clinical diagnosis of patients.In particular,for the lung diagnosis of covid-19 in 2020,Medical CT images are one of the important ways to confirm whether a patient’s lung is infected or has recovered.Although medical CT can provide powerful support for precision medicine and smart medicine,it is difficult to avoid certain harm to human body due to the X-ray radiation,and the risk of cancer caused by excessive absorption of X-ray.Therefore,it is a hot research topic that how to reduce the dose of X-ray as much as possible on the premise of ensuring the quality of CT images.The reduction of X-ray dose or CT scanning time tend to increase the CT image noise,thus affect the accuracy or reliability of clinical diagnosis.The study on CT image noise reduction work became a very key problem of image processing and analysis after CT image reconstruction.The research work of thesis based on the national natural science foundation of China youth project(project number: 61401049)and key industrial generic technology innovation in chongqing special project(project number: cstc2015zdcy-ztzx0072).During this period,some of research work finished in Shanghai First-imaging Information Technology co.,LTD.,a manufacturing medical CT equipment company.On the basis of extensive investigation and in-depth analysis of domestic and foreign literatures,this paper studied a method of image denoising for Cone Beam CT(CBCT)based on convolutional neural network.In this paper,through the comparative analysis of the basic principle,structure composition and imaging algorithm of spiral CT and CBCT,the sources and characteristics of CBCT image noise are identified.Based on this,this paper designs a CBCT image denoising network(PCDnet)based on convolutional neural network.In the selection of network framework,VGG network architecture without pooling layers is adopted,and its parameters are adjusted accordingly.It is effectively avoids the computational difficulty of classical algorithms such as BM3 D which cost a lot in the process of probability learning and inference.For the loss function,this paper adopts the feature similarity(FSIM)and mean square error(MSE).In order to verify the effectiveness of the network,a training data set including plaster head model CBCT images and seven groups of clinical human CBCT images was designed.Considering the limitation of computing resources,the data set is cut into 80×80 image blocks,and a popular method of data enhancement is adopted to expand the data set.The experimental results show that this model could achieve the identification of the overall characteristics of CT image and keep the details of the CT image processing without introducing a new noise.PCDnet noise reduction in both the subjective visual evaluation and PSNR,FSIM have significant advantages than classic noise reduction algorithms BM3 D,Dn CNN etc.objective evaluation indicators,such as to verify the effectiveness of the algorithm in this paper,the robustness and generalization ability.