Benign-Malignant Classification Algorithm For Pulmonary Nodules Based On CT Image

Lung cancer is a cancer with high morbidity and mortality.The benign-malignant classification of pulmonary nodules is the key to the early diagnosis of lung cancer.CT image is often used to screen malignant pulmonary nodules in clinic.However,due to the different sizes and shapes of pulmonary nodules,manual methods are prone to missed and mistaken detection.Studying the benign-malignant classification algorithms of pulmonary nodules can improve the accuracy of classification and help to increase the survival rate of lung cancer patients.Using deep learning method to achieve benign-malignant classification of pulmonary nodules.In the design of network structure,aiming at the characteristics of different sizes and shapes of pulmonary nodules,a multi-scale and multi-model integrated 3D convolution neural network architecture is proposed.It consists of three different basic network architectures: MSMME-VggNet,MSMME-ResNet and MSMME-InceResNet.Each basic network architecture contains three sub-networks,which process 3D images of 16*16*16,32*32*32 and 48*48*48 respectively.In the training of the network model,in order to solve the problem of small sample size in the lung cancer image data set,a two-stage data expansion method combining offline and online is proposed.Firstly,the data set is expanded offline by using the “random mask” method,which utilizing the similarity of the background around most lung nodules.Then,using the conventional image processing method for online expansion to improve the generalization ability of the network model.In the process of benign-malignant classification of pulmonary nodules,according to the diameter of pulmonary nodules,two subnetworks of each infrastructure network are selected for integrated prediction,and then the mean of prediction results of these three networks is used as the final benign-malignant classification results of pulmonary nodules.The proposed algorithm is compared with different methods on the published lung cancer dataset LIDC-IDRI.The accuracy rate is 89.43% and the AUC(Area Under Curve)is 0.9375.The experimental results are superior to other existing methods.It indicates that the proposed network structure,data preprocessing method and data expansion method are effective.