Optimization Algorithm For Coronary Artery Segmentation In CTA Images Based On Convolutional Neural Network

Coronary heart disease,as a common cardiovascular disease,seriously threatens people's health.In clinical practice,coronary Computed Tomography Angiography(CTA)is usually used to diagnose coronary artery-related diseases,but this technology will produce more image data.At present,these image data can only be processed mostly by manual or semi-automatic segmentation methods.The operation is too complicated and timeconsuming,and there are certain limitations.Therefore,it is particularly important to use automatic segmentation technology in clinical diagnosis.In recent years,Convolution Neural Network(CNN)has begun to be applied to medical image segmentation.The difference from traditional algorithms is that CNN can automatically extract the features of structures and patterns from a large amount of training data,and use these features for prediction to improve diagnosis efficiency.For the coronary CTA data,this paper studies the method of coronary artery segmentation using convolution neural network,including the selection of network structure,the optimization of network model and the post-processing of network segmentation results.The specific contents are as follows:(1)In the selection of network structure,this paper uses V-Net to process threedimensional coronary CTA images.Since traditional segmentation algorithms cannot meet the experimental needs in processing coronary artery segmentation,this paper mainly compares three commonly used three-dimensional image segmentation networks: 3D FCN,3D U-Net,and V-Net.In the evaluation process,performance metrics Jaccard Index(JI)and Dice Similarity Coefficient(DSC)are used as experimental evaluation criteria.The two coefficients in the V-Net segmentation results reach 0.7962 and 0.8843,respectively,which are better than the other two network results.Therefore,V-Net is selected as the segmentation network for coronary arteries in this paper.(2)In the optimization of network model,this paper incorporates Attention Gate(AG)model and Conditional Random Field-Recurrent Neural Network(CRF-RNN)model into V-Net at the same time to optimize the network model.On the one hand,the AG model is embedded in the network to suppress feature activation in irrelevant regions,highlighting the feature learning of the entire network for the vascular site;On the other hand,the CRFRNN model is embedded in the network to make the overall network have the properties of both V-Net and CRF,and can integrate CRF in V-Net to improve the performance of the network.The experimental results show that the optimized network merging the two models at the same time shows a good segmentation effect on coronary segmentation.The JI and DSC coefficients reach 0.8308 and 0.9062,respectively,which are better than the original V-Net network and the network integrated into a single model.(3)In the post-processing of network segmentation results,this paper uses the level set function to iteratively optimize the edges of blood vessels segmented by the network.Using deep learning to segment coronary arteries in CTA images,the edges of the segmented blood vessels are still rough.Therefore,this paper uses the traditional level set segmentation algorithm as a post-processing to improve this defect,and considers the result of network segmentation as the initial value of the level set function for optimization.The experimental results prove that the JI and DSC coefficients reach 0.8376 and 0.9103 respectively after adding the level set function,and the segmentation results of the optimized network model have been improved.