Application Of Deep Learning In Coronary Artery Segmentation

Coronary artery disease is one of the most lethal diseases worldwide.The diagnosis of this disease depends mainly on the patient’s coronary angiography.However,reading coronary angiograms requires the clinician must have a lot of experience and this process is very time-consuming.Therefore,an automatic segmentation technique of coronary angiography is urgently needed to reduce the burden of clinicians and improve the accuracy and efficiency of diagnosis.In recent years,with the great improvement of deep learning technology,its performance in image recognition has been greatly improved compared with traditional methods.Under this background,the research of medical image intelligent processing has been paid more and more attention,and some achievements have been made.Because CT angiography can produce a large amount of angiographic data,it is possible to use deep learning in the field of angiographic image segmentation.This project,combined with the latest deep learning technology,presents an end-to-end automatic segmentation algorithm for patients’coronary angiograms,which the name is PACNet(Pyramid Attention ConvLSTM Net).This method achieves an accuracy of 0.8913,a recall rate of 0.8206,and a Dice coefficient of 0.8508.This algorithm can provide clinicians with a real-time and accurate coronary angiography auxiliary segmentation system,and achieve the expected purpose of this project.The main work and innovations of this thesis are as follows:(1)A pyramid feature fusion module is proposed.The module can get a feature map with rich information by channel integration and size alignment of different scale feature maps in the process of up sampling.In coronary angiography,blood vessels are of different thickness,and the sensitivity of different scale feature maps to different thickness of blood vessels is different.Therefore,the pyramid structure enables the network to make full use of the background information near the pixels.(2)A ConvLSTM network module is proposed.LSTM is widely used in the field of natural language processing.This thesis modifies the structure appropriately,such as substituting the dot operation with convolution operation,so that it can be used in the field of compter vision.Using this module allows the network to progressively optimize the segmented results of an angiogram without giving the segmented results relatively once,as with other methods.(3)An attention weight module was proposed.This module is built to improve the ConvLSTM module.The Attention Module generates an attention weight matrix to weight the feature map which is the input to the ConvLSTM module.By doing this,in each time step,certain locations of the input signature map will be emphasized based on the current characteristics of the hidden cells.These locations are often more difficult and require multiple iterations.So that the network can learn more specifically.