Research On Metal Artifact Removal Method Of GT Image Based On Deep Learning

With the widespread application of computer tomography(CT)imaging technology in medical diagnosis,analysis and testing,and other fields,metal artifacts caused by metal implants such as metal dentures and pacemakers in patient's body lead the troubles of blurring or unclear boundary in reconstruction images.In order to enable doctors for more accurate observation and diagnosis with lesions and improve the quality of CT imaging,The problem of metal artifact reduction(MAR)has been widely studied and plenty of deep learning based MAR algorithms were proposed.However,existing deep learning approaches fail to fully analyze and extract the mask projection features.And the performance of transfer learning is not excellent enough,so that they do not have enough stability and strong robustness.In order to solve the problems aforementioned,this paper conducts researches on deep learning based MAR approaches,and propose three efficient artifact removal methods.First,a Dual Domain Network with Adaptive Mask and Weighted Mapping(D~2Net-AM-WM)is proposed to improve the ability of recognition and artifact removal.The design of adaptive mask design in sinogram domain network can automatically learn the appropriate mask projection coding to enrich the feature content extraction.After reconstruction,a weighted mapping module in image domain is designed to carry out a reasonable weight distribution for each encoding,so as to enhance the ability of identification on non-artifact information.Experimental results on synthetic dataset show that,D~2Net-AM-WM algorithm has 1.72dB and 0.95%gains in peak signal-to-noise ratio(PSNR)and structural similarity(SSIM)respectively when comparing with state-of-the-art method Du Do Net++.Second,a Reused Convolutional Network(RCN)is proposed to reduce the structural complexity of current unsupervised model.RCN transforms MAR problem to image generation problem.It utilizes two sub-networks to extract the artifact information and non-artifact information from CT images respectively,and generates corresponding artifact images and non-artifact images.In addition,RCN uses a two-stage reused method to achieve unsupervised learning,which also improves the model utilization efficiency,and simplifies overall workflow and model structure at the same time.Experimental results show that,RCN algorithm has enhanced the ability to reduce artifacts on synthetic dataset and clinical dataset,and has an exactly improvement of 0.14dB in PSNR and 0.55%in SSIM on synthetic dataset when comparing with the state-of-the-art method ADN.Finally,a Component Decomposition Network(Co De Net)is proposed to enhance feature learning and alleviate performance degradation in processes of MAR.Combining prior knowledge,Co De Net uses two sub-networks to decompose the tissue component and artifact component from CT images respectively,and a special attention-aggregation mechanism is designed globally and locally for the whole network.So that the model can better analyze the content information,and recover images with a way of unsupervised learning.Experimental results show that,Co De Net algorithm significantly enhances the ability of artifact reduction in both metrics and visualization on synthetic dataset and clinical dataset,and has a performance gain of 1.14 dB in PSNR and 1.3%in SSIM respectively when comparing with the state-of-the-art method ADN.