Research On Ctpa Pulmonary Embolism Images Segmentation Based On Deep Learning

Acute pulmonary embolism is a clinical syndrome of pulmonary circulatory dysfunction which is caused by endogenous or exogenous embolism blocking the main trunk of the pulmonary artery or its branch.The morbidity is second only to the diseases such as hypertension and it has third highest mortality rate.Due to its high morbidity and high risk,it has been more and more followed in recent years.Pulmonary embolism is diagnosed mainly through the technology of Pulmonary Angiography.Professional doctors view CTPA images and then manually segment the pulmonary embolism in the images.Professional doctors judge the degree of pulmonary embolism is depend on the segmentation results and the location of the pulmonary embolism.This process is very time-consuming and labor-intensive,and puts a lot of pressure on the doctor.Therefore,it has great significance to study the method of automatic segmentation the pulmonary embolism in CTPA images.In this work,Considered characteristics of pulmonary embolism images,we carry out related research on image segmentation based on deep learning and other related technologies.The main research contents are as follows:(1)A pulmonary embolism image segmentation method has been proposed by fully convolutional neural network.In this method,Dice Loss is introduced to solve the problem of imbalance classification of the training dataset.After analyzing the differences between the network structure characteristics and segmentation accuracy about FCN-32s,FCN-16s and FCN-8s,the methods for improving network segmentation precision has been discussed,which laid the foundation for the network improvements.(2)In this part,We carry out a method for pulmonary embolism images segmentation based on U-net.Combined with the basic theory of transfer learning,and the model migration method is used to effectively improve the U-net segmentation accuracy.At the same time,in order to overcome the problem of unstable training of Dice Loss,a new loss function is proposed in combination with the design idea of Focal Loss and related theories,the effectiveness of transfer learning and the new loss function is proved through the comparative analysis by experiments.(3)After studying the basic structure and principles of U-net,the intrinsic nature of U-net segmentation accuracy better than Fully Convolutional Neural network has been deeply analyzed.Combining the ideas and basic principles of ResNet and DenseNet,an improvement U-net structure has been proposed.The improved U-net builds a powerful feature extraction network with the basic structure of the residual module.And using the intermediate feature fusion module named Concat Block,it retains more intermediate features of the contraction path,and merges the fused features into the expansion path of U-net.The experimental results prove the effectiveness of the improvement.(4)Considering the needs of the actual scene about Pulmonary embolism risk assessment,a method of pulmonary embolism images segmentation based on Mask RCNN is implemented,which can give the results of pulmonary embolism segmentation and locating the location of the pulmonary embolism.On the basis of this work,the Mask RCNN is improved by using Group convolution,which reduces the number of network parameters and optimizes the inference speed of the network.The improved method can reduce the GPU memory and time occupation more than thirty percent,and the final accuracy is almost constant.