2D/3D CT Reconstruction From Incomplete Projections By Using A Contextual Autoencoder Network

As a routine and effective medical diagnostic tool,X-ray computed tomography(CT)provides clear images of human anatomy for clinical diagnosis.But one complete CT scan usually produces a higher degree of ionizing radiation.In a cone beam computed tomography(CBCT)scan,the patient may be exposed to four times as much radiation as a conventional CT scan.Although reducing the CT scan views will reduce the ionizing radiation effectively,it may result in partial loss of the projection data at some scanning views.Eventually,the reconstructed CT images are degraded.How to reconstruct a clinically satisfactory CT image from incomplete CT projection data has important social value and research value.In order to improve reconstructed CT image with limited-view CBCT projections data,this paper proposes a limited-view CBCT reconstruction algorithm based on contextual autoencoder network.The proposed method includes data preprocessing,training stage and testing stage.Since the contextual autoencoder network used in this paper can only deal with two-dimensional data,it is necessary to use the data preprocessing stage to slice the CBCT three-dimensional projection data into a series of two-dimensional data,which will be treated as the input data of the training stage.The network structure in the training stage employs a generative adversarial network(GAN)to estimate the missing part of CBCT projections.The generative network has similar structure to autoencoder and is composed of encoder part and decoder part.It is constructed by using a series of convolutional neural networks(CNN).The encoder part aims to learn the structural information of the incomplete two-dimensional data,and then the decoder part generates the missing parts of the two-dimensional data.Then,the discriminant network,whose structure is simiar to the encoder part of the generative network,judges whether the generated two-dimensional data of the network estimation is true or not.The trained generation network estimates the missing part of the newly preprocessed two-dimensional data,and then stacks the completed two-dimensional data in order to obtain the full-view threedimensional CBCT projection data.Finally,the Feldkamp algorithm will be used to reconstruct CT images from completed CBCT 3D projection data.The experimental results show that the proposed method can improve the reconstructed CBCT images effectively when 3D CBCT projection data in some scan views are missing.Secondly,in order to improve reconstructed CT image with truncated two-dimensional CT projection data,this paper proposes a CT image reconstruction algorithm with truncated projections estimated by a contextual autoencoder network.The proposed algorithm includes a training stage and a testing stage.In the training stage,the obtained truncated projection data is used as an input to train the contextual autoencoder network to estimate the missing parts of truncated data.In the test stage,as the new truncated projection data is input to the trained generative network,the generative network will produce the missing parts of the new truncated projection data.After combining the generated projection data and the known ones,the proposed method obtain the completed CT projections.Finally,the CT images are reconstructed from the completed projection data through the filtered backprojection method.The experimental results show that the proposed algorithm can improve the quality of CT image reconstruction effectively when some CT projection data in detector direction are truncated.This paper uses the contextual autoencoder network to improve the reconstructed CT image in the absence of partial projection data.It can achieve clinically satisfactory CT images with reduced doses of X-rays.