Research On Lung Parenchyma Segmentation And Lung Nodule Detection Based On Deep Convolution Neural Network

With the improvement of medical image technology,the computer-aided detection algorithm based on medical image has made a lot of contributions in assisting doctors to diagnose the disease area.The analysis and detection of lung medical images are often used in the field of medical and health services to assist doctors in the diagnosis of lung diseases.Computer-aided detection is helpful for radiologists to find missing pulmonary nodules through image interpretation in lung CT screening.However,the existing methods still have many limitations in medical image analysis.In order to improve the performance of pulmonary nodule detection,this thesis studies several key problems in lung parenchyma segmentation and pulmonary nodule detection based on deep convolution neural network.The main work is as follows:(1)In lung parenchyma segmentation,the edge information of lung CT image is often fuzzy,but the current research methods mostly ignore the feature extraction and utilization of lung parenchyma edge contour,which is very unfavorable for the application of deep convolution neural network in lung parenchyma segmentation.First of all,in view of the problems of neglecting edge features and low segmentation accuracy in the lung parenchyma segmentation model,this thesis proposes an Attention Gabor Network(AGNet)based on the deep convolution neural network for medical image segmentation,which can automatically pay more attention to the target contour and continuously improve the segmentation performance.Secondly,aiming at the problem that the loss function of segmented network can not fully adapt to attention mechanism,the loss function is optimized and multi task loss function is used.The loss of common features and edge features are fused according to certain weight.Then,aiming at the over fitting problem of the network,a drop strategy is used.According to a certain proportion of random selection of relevant features,it can effectively alleviate the over fitting problem and achieve an effective balance between calculation efficiency and accuracy.In addition,this thesis also uses the open data set for experiments to verify the superiority of the model.(2)In pulmonary nodule recognition,firstly,in order to improve the detection performance of the model,a multi-scale deep convolution detection network based on attention mechanism and feature pyramid network is designed.This network takes the deep convolution neural network as the feature extraction structure,and uses the feature pyramid network structure to realize the attention mechanism,so as 1to fuse the features of different scales and strengthen the recognition ability of small targets.At the same time,in order to further optimize the ability of accurate regression of boundary box position in the network,the loss function was redesigned by introducing GIo U loss to further improve thedetection efficiency of pulmonary nodules.Then,through the experiment on the lung nodule image data set,it is verified that the designed network can improve the accuracy and efficiency of the lung nodule detection model,and the detection results are visualized in order to better show the network performance.