Research On Deep Learning Based Low-dose CT Image Reconstruction Method

Computed tomography(CT)plays an important role in imaging diagnosis,but radiation dose exposure caused by the detection process is also a potential problem worthy of our consideration.In medical diagnosis,in order to reduce the risk of radiation exposure,usually adopt the way of reducing pipe current or voltage to achieve the reduction of radiation dose,this way can avoid too much of a patient to X-ray radiation,but reduce the radiation dose in reducing radiation exposure risk at the same time,also improve the level of the noise,which makes the reconstruction results to increase produce streak artifacts,The constructed image will also degrade seriously due to the presence of noise and artifacts,which makes it difficult to detect small changes in the internal organization of the patient,affecting the image quality and the final diagnostic results,and greatly weakening its diagnostic performance.Studies have shown that image reconstruction of low-dose CT images related to radiation dose has become a very concerned issue in the medical community.In this paper,an in-depth study was conducted on how to balance the quality of CT images with radiation dose.For low-dose CT image reconstruction methods,the following two aspects were mainly studied.(1)In view of the fact that reducing radiation dose may lead to the increase of noise and artifact,the method of eliminating noise and artifact of LDCT image has attracted more and more attention.In this paper,a low dose CT image reconstruction method based on generative adversarial network(GAN)is proposed to solve the problems of noise residue,excessive smooth structure and false damage caused by noise in CT reconstruction results.Due to the wide application of U-NET in medical imaging,this paper takes the improved U-NET as the first stage of generator network,and uses generative adversarial network to denoise the sinusoidal image,which can reduce the image artifact caused by the reduction of dose.In this process,the residual U-NET network is used to realize the mapping from the CT projection data with strong noise to the projected images with high signal-to-noise ratio,and the gradient dispersion problem caused by deepening the network layer number can be avoided by using the residual network.The original CT projection image with noise artifacts is directly used as the input of the network.In the case of high noise of LDCT projection data,the FBP reconstructed LDCT image will be affected by fringe artifacts and noise.Therefore,in the second stage of the generator,multi-scale extraction blocks are used to enhance the restoration of texture details.GAN focuses on the statistical migration of data noise distribution from strong to weak,so the VGG network is used to improve high visual sensitivity and suppress noise by comparing with the perceptual characteristics in the established feature space of the original image.Experimental results on Mayo dataset show that the proposed method has better visual effect than direct post-processing image reconstruction.(2)Due to the discomfort of image reconstruction,high quality image reconstruction based on low-dose CT data to improve the diagnostic performance is a challenging problem.In order to better adapt to the large number of image data in imaging diagnosis,this paper designed a dual-path multi-feature subnetwork for low-dose CT image post-processing and reconstruction.This method is based on two parallel subnetworks for feature extraction.The features extracted from the two subnetworks can get acceptance domains of different sizes.Concat is then used to aggregate the features captured on the two paths,which can not only restore the organizational structure more fully,but also improve the robustness and generalization ability of the network.The use of void convolution can improve the perceptual field by using the context information of the image domain,while the use of sub-pixel convolution is more conducive to alleviate the loss of pixels,gather multi-scale features in a single path,and provide more abundant information for image reconstruction,thus providing image quality.Experimental results show that the method proposed in this paper converges quickly and can simplify the training difficulty.At the same time,with the deepening of the network,more image details can be expressed.Compared with the traditional method,the performance of this method in restoring the texture details of medical images has been improved.