Left Ventricular Segmentation On Ultrasound Images Using Deep Layer Aggregation For Residual Dense Network

In recent years,the number of people suffering from cardiovascular diseases has increased,and cardiovascular diseases have seriously threatened human life.In order to allow patients sufficient treatment time,doctors should diagnose cardiovascular diseases for the patients as early as possible.The segmentation results of the left ventricle ultrasound images can assist doctors in the diagnosis of cardiovascular diseases,but the left ventricular ultrasound images have the characteristics of strong noise,weak edges,and complicated ventricular structures,which not only make the segmentation of the left ventricular ultrasound images quite difficult,but also lead to the poor efficiency and accuracy of segmentation results.In order to improve the accuracy and speed of the image segmentation results,the left ventricular segmentation method on ultrasound images using deep layer aggregation for residual dense network is proposed in this thesis,whose research content includes image preprocessing,design of segmentation network and implementation of segmentation system.(1)In order to reduce the influence of the tissues around the left ventricle(fat or lung,etc.)on the segmentation results of the ultrasound images,in this thesis,the threshold segmentation method is used to crop the images directionally.Firstly,the pixels of the left ventricular ultrasound images are counted,and the images are binarized according to the thresholds obtained by the statistics,which is used to distinguish the backgrounds,left ventricles,lungs,and fat areas in the left ventricular ultrasound images.Secondly,morphological processes(open operation,erosion,close operation)are used to delete the disturbing blocks of the images and fill their closed areas,so that the adjacent areas are connected together to reduce the number of image contours in the images.Thirdly,the target contours are locked according to the prior information of the left ventricular ultrasound images,and the left ventricular ultrasound images are cropped according to the locked target ranges.(2)In order to improve the segmentation accuracy of the left ventricle ultrasound images,based on the encoding-decoding framework neural network,the residual dense network is designed for left ventricle ultrasound image segmentation in this thesis.Firstly,combining the advantages of residual network and dense network,the residual dense block is designed for image feature extraction,and it is applied to the up-sampling and down-sampling channels of the encoding-decoding neural network.Secondly,the depth of the segmentation network is established through experiments.Thirdly,the segmentation masks are post-processed by using the fully connected random field.Experimental results show that the proposed algorithm has a good segmentation performance on the test dataset.(3)In order to improve the segmentation accuracy and generalization performance of the residual dense network,the left ventricular segmentation method on ultrasound images using deep layer aggregation for residual dense network is proposed in this thesis.Firstly,based on the dense residual segmentation network,the down-sampling channel is designed to extract image features.Secondly,in order to make the shallow and deep feature information of images more closely integrated,we adopt a network connection method called deep layer aggregation.Thirdly,we optimize the neural network.For the redundant part of the network,we use the deep supervision method to pruning the network.Consequently,we simplify the network structure and improve the speed and generalization performance of the neural network.After these steps,we have completed the design of the segmentation network.The experimental results on the test dataset show that the average accuracy of the segmentation result is 95.68%,the average cross ratio is 97.13%,Dice is 97.15%,the average vertical distance is 0.31 mm,and the contour yield is 99.32%.Compared with the current six segmentation popular algorithms,the proposed algorithm achieves higher segmentation precision of the left ventricle on the ultrasound images.(4)In order to promote the application of the left ventricle ultrasound image segmentation algorithm in real life,based on the segmentation algorithms mentioned in this thesis,the left ventricle ultrasound image segmentation system was designed using the Pycharm under the Ubuntu operating system.The system extracts 21 ultrasound images for segmentation within one cycle of heart beating.The segmentation results have high accuracy and contour yield,which verifies the efficiency of the algorithms.