Research On Filter Learning Network Based Optimization Of Cone-beam CT Analytic Reconstruction

Traditional reconstruction algorithms of Cone-beam Computed Tomography(CBCT)are mainly divided into Analytic Reconstruction(AR)and Iterative Reconstruction(IR).FDK algorithm is the most classic AR algorithm,which is simple and well-explained.As a filtered backprojection algorithm,filter operation is the key factor affecting the image quality of FDK reconstruction.IR algorithms mainly include algebraic iterative reconstruction and statistical iterative reconstruction,which requires multiple projection and backprojection operations to reconstruct high-quality images.Thus,IR algorithms are usually complicated and time-consuming.In recent years,neural network models have been widely used in the field of CBCT image reconstruction.For a large number of parameters and high complexity in the convolutional layer and fully-connected layer,many models require high storage and computing power,causing limited generalization and practicability.In this thesis,an FDK algorithm-based filter learning network model is proposed,which effectively combines the advantages of fast reconstruction and learnable network method,using high-quality paradigm data to guide the optimization of filter weights.Firstly,based on the linear equation describing CBCT imaging,the matrix operation expression of the reconstructed object is derived and matched with steps of FDK reconstruction.Thus,the reconstruction process is discretized into a cosine-weighted operator,filter operator,and backprojection operator.Then,the cosine-weighted operator is realized as a multiplication layer of a two-dimensional matrix,whose elements are the cosine function of the incident angle.The filter operator is realized as a one-dimensional filter multiplication layer in the Fourier frequency domain and initialized as a ramp function.The backprojection operator comes from the packaging and compilation of the backprojection reconstruction function and the neural network layer is constructed in the form of a dynamic link library function.Finally,the above neural network layers are cascaded with an activation function to build the filter learning model based on the FDK reconstruction.By setting the filter weights as the only trainable parameters,the filter is optimized during the training process.In this thesis,the iterative reconstruction images are used as the paradigm data to guide the model to optimize the filter.We construct the study as the following four parts:(1)The images reconstructed by the FDK algorithm with a specific filter function are used to train the model,and the consistency between the learned filter and the target filter is explored to verify the feasibility of the proposed method.(2)The optimal iterative reconstructed images are used to train the model and verify whether the learned filter can effectively improve the quality of FDK reconstruction.(3)The oversmooth iterative images are used to train the model and explore the effectiveness of optimizing filter weights with different feature paradigms.(4)Two reconstruction strategies of U-net post-processing and backprojected filtering are used for comparative analysis.The experimental results demonstrate that the filter learning network performs well at optimizing filter weights,and the learned filter keeps high consistency with the target filter.The FDK reconstruction with the filter learned from iterative paradigm images gets a better image quality and resolution than standard FDK,and the structural similarity and peak signal-to-noise ratio are improved by 13.75%and 42.78%,respectively.In addition,the filter learning network holds an equal reconstruction speed with standard FDK,which is nearly 42 times higher than iterative reconstruction,then,the stability and generalization are obviously better than both post-processing and backprojected filtering strategies.