Low Dose CT Image Denoising Based On Generative Adversarial Network

With the widespread use of medical CT,many patients benefit greatly from CT scanning.However,excessive X-ray radiation can increase the risk of cancer and genetic diseases in patients.In clinical medicine,radiation should be reduced as much as possible while ensuring acceptable diagnostic accuracy,so low-dose CT scan came into being.However,the reduction of radiation dose often leads to the increase of speckle noise and non-stationary fringe artifacts in reconstructed images,resulting in the degradation of image quality and affecting clinical diagnosis.In order to improve image quality,low dose CT image denoising has become an important research direction in the field of medical imaging.Many algorithms have been developed to improve low-dose CT images.In this paper,the application of generative adversarial network in low-dose CT image denoising is studied.Based on generative adversarial network,three low-dose CT image denoising methods are proposed.The main research work and achievements are as follows:(1)A low-dose CT image denoising method based on subspace projection is proposed.Aiming at the low efficiency of traditional convolution methods in modeling the global structural information in CT images,a generative adversarial network denoising model is designed,which uses a subspace projection denoising network as a generator,and is trained against a discriminator.The projection network reconstructs the noisy image in the subspace through the basis vector generation and projection operation,and the projection operation naturally introduces the global structure information into the denoising process.To make the reconstructed image clearer and more realistic,the network is jointly optimized using adversarial loss and traditional loss.A simple and efficient selfsupervised learning method for denoising synthetic noise is further proposed,which not only increases the robustness of the model but also accelerates the convergence speed through representation learning.Experimental results on public CT datasets show that the proposed low-dose CT image denoising algorithm can effectively suppress noise while retaining clear edges and structures,reducing image blur.For weak texture areas,compared with conventional convolution methods have better repair effect.(2)A dual adversarial networks denoising model combined with perceptual loss is proposed to solve the problem of CT image denoising.The dual adversarial network consists of two generators,which model the joint distribution of clean-noise images from the perspective of image denoising and noise generation.The denoising and noise generator jointly learn and guide each other to achieve better denoising effect.Moreover,a RES-UNET residual learning network is designed to achieve denoising and noise generation tasks.The introduction of residual blocks can make the network retain more image details.At the same time,in order to make the denoising effect more consistent with human visual cognition,a weighted mixed loss function is proposed to optimize the network,which includes the mixed loss of confrontation loss,perception loss and minimum absolute deviation loss.In this process,in order to better calculate the perception loss,a mask self-supervision method is proposed to train a perception loss model for CT image domain.Experimental results show that the proposed denoising model can suppress the noise and retain the edge contour and texture details of low-dose CT images.(3)A low dose CT image denoising algorithm based on unsupervised image domain transform is proposed.For unpaired CT data sets,the supervised method cannot train an effective denoising model.However,the current unsupervised CT image denoising methods have not achieved good results.Therefore,an unsupervised low dose CT image denoising method is proposed based on domain transform framework.Denoising is achieved by dissociation representation learning.Specifically,the clear image information and the noise information in low-dose CT images are represented separately by the dissociation representation,and the joint distribution is modeled by the cross transform,and the clear and noise-free image is reconstructed.At the same time,in order to ensure that details and background offset will not be lost after the transformation from low-dose CT images to normal-dose CT images,cyclic consistency constraints,adversarial domain constraints and background semantic consistency constraints are used to optimize the network.Experimental results show that compared with other unsupervised algorithms,the proposed algorithm can suppress noise and bar artifacts more effectively,and the denoised image is more consistent with the real normal dose image.