Research On Sparse-View CT Reconstruction Algorithm Based On Deep Learning

X-ray CT imaging technology can accurately image the internal anatomical structure of patients and has been widely used in clinical medicine.However,radiation generated by CT scanning will cause certain harm to the human body,and too high a radiation dose will increase the risk of cancer in the human body.At present,one of the commonly used solutions to reduce the radiation dose is to reduce the scanning views and adopt sparse-view scanning.However,the projection data obtained from the sparse-view is not complete,and the inverse solution of the mathematical model leads to serious strip artifacts in the reconstructed image.Therefore,model-driven iterative reconstruction algorithms are usually used for the sparse-view CT reconstruction.However,model-driven iterative reconstruction usually requires designing different regularization terms,making appropriate prior assumptions for different types of CT images,and involves tedious manual parameter tuning.In recent years,with the development of deep learning technology,scholars have proposed data-driven CT reconstruction methods,which can realize the adaptive training of parameters,and the model has stronger adaptability.But the deep learning methods usually relies on large sample data,in the field of clinical medicine,because for the patient’s privacy protection,data samples are less,and the depth of the existing neural network model can easily lead to data fitting,and due to large differences between each part of the human body,directly end-to-end network model generalization ability is weak and poor robustness.Therefore,in view of the problems existing in the existing deep learning-based reconstruction methods,this paper combines model-driven and data-driven methods and constrains the network through an iterative algorithm framework to reduce the dependence on data samples and improve the generalization ability of the network.The main research work is as follows:(1)In the field of clinical medicine,it is difficult to obtain patient data samples,and the scale of data sets is small.To address the above problems,a model-driven and data-driven sparse-view CT reconstruction algorithm is proposed.The algorithm adopts the deep learning strategy,and expands the ADMM iterative framework for different levels of the network,using CNN as the a priori model,on the one hand,can learn adaptive iteration prior items and related parameters in the model,to avoid the artificial prior information of design and onerous parameter tuning,on the other hand,make the network is governed by the iterative model framework,reduce the strong dependence on data samples.Experimental results indicate that the proposed algorithm can availably suppress artifacts and noise under the condition of a relatively small dataset.,and enhance the quality of reconstructed images.(2)For the problem that the existing CNN-based image reconstruction algorithms cannot model long-distance pixel correlation due to the limited receptive field of CNN,a CT reconstruction method ground on the Transformer model is proposed.By introducing an atention mchanism to expand the receptive field range,the global features of images are learned.At the same time,the MDTA(Multi-Dconv head Transposed Attention)model and GDFN(Gated-Dconv Feed-forward Network)are used to decrease the computational complexity of the network and focus on image details that are complementary to other layers.The experimental results show that the algorithm can further preserve the details of the reconstructed image and increase the precision of the reconstructed image.(3)The reconstruction algorithm based on iteratively unrolling network has high reconstruction accuracy,but the number of network layers is deep,and has the weaknesses of low training efficiency and slow reconstruction speed.The reconstruction network based on domain-transform has high training efficiency,but poor reconstruction accuracy and insufficient generalization ability.Aming to solve thees problems,the CT reconstruction method with coupled iterative-domain transform is implemented.The domain-transform reconstruction parallel branch is used to gain the computational efficiency of the network,and the basic reconstructed image is obtained by directly learning the popular mapping from the projection domain to the image-domain;The iterative reconstruction branch is used to constrain the training and reconstruction process of the network and prevent the network from overfitting to improve the generalization ability and stability of the algorithm.Finally,the output of the two branches is fused through a fusion model based on an adaptive attention mechanism to form the final reconstructed image.At the same time,based on the sparse non-negative matrix factorization theory,the domain-transform network AUTOMAP is improved in a targeted manner,which largely diminish a great deal of parameters of the fully connected layer in the model.The experimental results suggest that the algorithm can meet the reconstruction accuracy better,and has a certain generalization ability.