Low-Dose CT Image Processing And Reconstruction Using Deep Residual Network

The wide application of X-ray computed tomography(Computed Tomography,CT)makes low-dose CT become the main direction of clinical application research.However,reducing the radiation dose in CT will obviously increase the noise and artifacts in the image,which will significantly reduce the accuracy of the diagnosis of clinical radiologists.The low-dose CT processing methods which are widely used in recent years include projection domain preprocessing,iterative reconstruction,image post-processing and deep learning methods.The main goal of these methods is to be close to or even reach the image quality of routine dose scanning CT under low-dose scanning conditions.Conventional low-dose CT processing and reconstruction techniques will easily lead to feature loss of the tissue details.In this paper,based on deep learning methods and convolutional neural network,the complex problem of low dose noise-artifact prediction is modeled and fitted by constructing and training deep feature representation network.The aim of this paper is to improve the quality of low dose CT imaging.Reduce the exposure of the patient to scanning radiation,and provide effective assistance to clinicians in the detection and diagnosis.This paper is divided into two parts:In this paper,we make research on replacing conventional fixed kernels in filtered backprojection using adaptive convolution filter network instead of filter backprojection.The AFDenseNet uses adaptive filtering elements to filter the projection data instead of fixed filter kernels in FBP,and convolutional opreation is carried out both in frequency domain and space domain at the same time.The contrast and high frequency details of the image can be saved to the maximum extent while filtering the noise.AF-DenseNet has the characteristics of fast computation speed of the filtered back projection algorithm,and the image quality of which is close to the iterative reconstruction methods at the same time.This network is used to solve the disadvantages of end-to-end deep learning image processing methods.The end-to-end image process is combined with the domain transformation reconstruction process.A Domain Progressive Residual Network(DP-ResNet)processing method combining projection domain pre-processing and image domain post-processing is presented in this paper.The network structure includes projection domain pre-processing sub-network(SD-Net),filtered backprojection reconstruction(FBP)and image domain post-processing sub-network(ID-Net).The main idea of this algorithm is: firstly,the low-dose projection data is preprocessed by SD-Net,the image details are preserved to the maximum extent,and the image noise is reduced to a certain extent.The filtered back projection is used to reconstruct the image data to be processed in next step.Finally,ID-Net is used to post-process the image data to further reduce noise-artifact and restore the occluded tissue details.Both the simulated data from the AAPM Grand Challenge and the real low-dose projection data from the UIH company show that the DP-ResNet deep learning network combines two network processing methods in different fields and achieves a significant improvement in imaging performance.