Removal Of Medical Motion Artifact Based On Generative Adversarial Networks

Motion artifact in Magnetic Resonance Imaging(MRI)refers to the blur area or ghost in MRI,due to the autonomous or physiological movement of the scanned part.This is also common unavoidable artifact in medical artifact.In clinical diagnosis,motion artifact in MRI will affect the doctor’s diagnosis.Although reacquisition can get MRI without motion artifact,it will cause additional time costs and economic cost.Therefore,the removal of motion artifact by image processing is a meaningful work.In recent years,generative adversarial networks(GANs)have achieved excellent results in natural image restoration.But for motion artifact removal in MRI,because of the complex environment of magnetic resonance imaging,the motion artifact removal model based on GANs make further improvement.In this thesis,for the relatively simple semantic information of medical images,deep neural networks are easy to over fit,and the existing methods have not paid attention to the local consistency and domain consistency of MRI,two motion artifact removal models for MRI are proposed based on GANs.The specific research work is as follows:1.Aiming at the problem that the current deep learning-based motion artifact removal models have not pay attention to the local consistency of MRI,this thesis proposes a motion artifact removal model that can maintain the local consistency of the repaired image implemented by local consistency loss.In addition,because medical images have less semantic information,in order to make full use of the feature of different levels of the image to get high-quality repaired images,residual structure is introduced into the model to extract high-level features of the image,and skip-connection is introduced into the model to merge low-level feature and high-level feature of the image.Experiments show that the model makes a certain improvement in simulated MRI datasets and real motion artifact data on objective evaluation metrics,and the local texture of the low-contrast area in the repaired image is more similar to the texture of the label image.2.Aiming at the problem that the phase shift of frequency domain during MR acquisition directly causes motion artifacts in spatial domain and the current deep learning-based motion artifact removal models have not pay attention to the domain consistency of MRI,a joint optimization model based on frequency domain and spatial domain is proposed to remove motion artifact in MRI.This model converts the original MRI with motion artifact into K-space data(frequency domain),and optimizes the model with K-space data and spatial data jointly to maintain the domain consistency.Experiments show that the model can produce realistic image textures in both highcontrast areas and low-contrast areas.3.Design and implement a motion artifact removal system.The system encapsulates the two motion artifact removal models proposed in this thesis,and users can choose different motion artifact removal models according to their requirements.The system implements the single-shot motion artifact removal function and the batch motion artifact removal function to meet different requirements for real-time observation of single instance and batch motion artifact removal.The motion artifact removal system provides a friendly graphical operation experimental platform for subsequent experiments,and provides convenience for subsequent experimental research.