Motion Artifact Removal In Medical Images Based On Deep Learning

Medical image quality is one of the key factors affecting clinical diagnosis.However,the diversity of the image acquisition protocols,receiver coils,organ density and anatomical regions often leads to various noise and artifacts,with MRI motion artifacts being the most represented one and difficult to avoid.Therefore,how to suppress and remove MRI motion artifacts has been widely concerned in medical imaging processing.Recently,more and more deep-learning(DL)-based methods have been proposed for this task,which mostly based on MR images with simulated artifacts for training.These methods have shown remarkable performance of removing the artifacts by learning features in the images,and guide model optimization through loss functions.We address the limitation of existing DL-based MRI motion artifact removal models,which overlook the information between medical image sequences and the local information.Existing methods also neglect the frequency domain data of MR images and the poor generalization performance on removing real artifacts.Two medical image motion artifact removal models and a unsupervised domain adaptation(UDA)learning framework for cross-domain real artifacts removal are proposed to improve upon these issues.The content can be divided into three parts:1.A motion artifact removal model based on generative adversarial network is proposed by fusing low-level,high-level,and contextual features of medical images.The model uses skip-connection residual networks to combine high-level and low-level features to solve the problem of insufficient semantic information in medical images.It also utilizes contextual features between adjacent slices to complement the anatomical details and content information.Taking advantage of the proposed region consistency loss,the network,trained by jointly optimizing all the loss functions,removes artifacts while maintaining the texture structure of the generated images.Experimental results show that compared to other motion artifact removal models,the proposed model performs better in both artifact removal and structure preservation.2.Based on the characteristics of MR images,a motion artifact removal model that fuses frequency and spatial domains is proposed.The complex convolution generators are introduced to process frequency domain data to fully utilize the unique information of K-space in MR images.A data consistency loss between frequency and spatial domains is proposed to guide the training stage,which avoid the deviation that may arise from training in a single domain by synthesizing the characteristics of both domains.The experiments demonstrate that fusing frequency and spatial domain data of MR images can make the model more robust in removing motion artifacts.3.An UDA cross-domain real artifacts removal network is proposed.An unsupervised self-training paradigm is adopted.By generating clean pseudo-labels for real artifact images through an uncertainty mask and training the network with both real and simulated artifact images,this network successfully solves the domain shift problem between simulated artifact datasets(source domain)and real artifact datasets(target domain).Experimental results show that compared to other supervised learning methods,this network perform better on real artifacts removal.