Left Atrium Segmentation And Quantitative Analysis Of Atrial Wall Fibrosis In Delayed-Enhancement MRI Images

Atrial fibrillation(AF)is a common arrhythmia disease in clinic of which pathogenesis is closely related to left atrial wall fibrosis.Delayedenhancement can clearly describe the area of atrial wall and the distribution of myocardial fibrosis,so it has been widely used in atrial fibrillation ablation.The accurate segmentation of left atrium and quantitative analysis of atrial wall fibrosis in delayed-enhancement MRI images can provide important diagnostic and therapeutic reference for ablation surgery.The work of this paper is summarized as follows:Firstly,a left atrial segmentation method based on Gaussian mixture model and level set is proposed.In the unlabeled clinical data,the robustness is improved by heart-center based image enhancement method,and the segmentation accuracy is improved through combination of regional information and edge information in level set.The experimental results show that the method can obtain accurate segmentation of the left atrium on 55 layers of slice without artificial intervention in totally 96 slices of four clinical cases containing left atrium.Secondly,a segmentation method combining convolution neural network and cyclic neural network is proposed.On the labeled public data,through improvement in model capacity,model structures and loss function,make the method more suitable in left atrium segmentation.The experimental results show that on each 20 test images of LASC and 2018 Atrial Segmentation Challenge,this method can achieve average Dice coefficients over 0.9,and the computing time of each test image is less than 30 seconds.Furthermore,a semi-supervised learning method is proposed based on the level set method and deep learning.Pre-trained neural network model is accomplished by a small amount of labeled data,and the prediction result of the model is used as the initialization of the level set on the unlabeled data,and then the evolution result of the level set is used as the optimal supervisory information to guide the updating of the neural network,so as to improve its segmentation effect and generalization ability.The experimental results show that this method can increase the average Dice coefficient of the test set by about one percentage through the introduction of unlabeled data on the 2018 Atrial Segmentation Challenge.Finally,a quantitative segmentation method of atrial wall fibrosis tissue based on graph cut algorithm and connectivity analysis is proposed.On the basis of left atrium segmentation,the area of atrial wall is determined.Gray scale and morphological characteristics of fibrous tissue are used to construct a graph model.Connectivity analysis is used as post-processing to improve the accuracy of segmentation,and the proportion of fibrous tissue in the atrial wall is calculated as a quantitative index.The experimental results show that on the basis of accurate left atrial segmentation,the framework can segment and quantify fibrotic tissue accurately,and the quantified index are basically consistent with the clinical diagnosis after evaluation by clinicians.