| In recent years,cardiovascular disease has become one of the major causes of increased mortality worldwide.The rapid development of medical imaging technology has made great contributions to improving the accuracy of cardiovascular disease prediction.At present,Coronary CT Angiography(CCTA)has played an important role in the diagnosis and research of Coronary artery diseases.However,during CT imaging,the chest and heart are more susceptible to undesirable breath-holding and heart beats than other body parts,resulting in motion artifacts in CCTA images.Motion artifacts can cause image distortion such as blur,double or severe deformation of internal structure,which will affect the accuracy of diagnosis.Therefore,quantifying the artifacts in the images before the imaging doctors read the images can reduce misjudgment and obtain objective evaluation of image quality,and it is particularly important for establishing automatic computer-aided diagnosis.In CCTA images,the motion artifact in the Right Coronary Artery(RCA)is the most significant and can easily affect the subsequent diagnosis,so this paper mainly carried out the detection and quantitative study of motion artifact in the ROI region of 60×60 at the RCA.(1)Motion track-based simulation of CCTA image artifacts: In order to solve the problem of lack of quantitative annotation data set in the evaluation of motion artifacts in CCTA images,this paper proposes a motion artifact simulation algorithm.The breathing motion and heart beating caused by the moving artifacts at RCA were modeled,and the motion trajectory was generated by their interaction.After that,the motion trajectory was up-sampled,and the motion blur kernel obtained was convolved with the CCTA image without motion artifacts to generate the image containing motion artifacts.Finally,the PSNR and structural similarity coefficient are used to quantitatively mark the fraction of the images with moving artifacts as the data set of this paper.(2)Artifact evaluation model based on machine learning: the existing algorithms mostly study artifact distortion in natural images.Considering the imaging differences between medical images and natural images,the image data set used in the experiment was analyzed in this paper.According to the characteristics of artifacts in CCTA images,local binary mode,Scharr filter and Gabor filter were selected to extract image texture features.The extracted texture features and the normalized brightness features of CCTA images are used as inputs to the Support Vector Regression(SVR)model,and the SVR is used to learn the nonlinear mapping of features to predicted scores.Finally,the proposed algorithm is tested on the data set in this paper and the public data set respectively,and the results show that the proposed algorithm is superior to other algorithms.(3)Deep learning-based artifact evaluation model: Aiming at the problem that the data set in this paper is small and the background in most CCTA images is dark,this paper selects Resnet50 as the feature extraction network,and integrates the attention mechanism to improve the model’s performance of artifact evaluation.In addition,the content understanding hyper network is added to the traditional prediction network.It can generate different parameters for the quantitative prediction module of artifacts according to the different input images,which increases the adaptability of the network model.In this paper,the network model with attention mechanism and the network model without attention mechanism are experimented on the data set of this paper and the public data set respectively,and the results confirm the performance of the proposed algorithm. |
