Research On Classification Method Of Benign And Malignant Pulmonary Nodules Based On Convolutional Neural Network

Computer-aided diagnosis technology is one of the hotspots in current scientific research.The classification of benign and malignant pulmonary nodules in CT images is part of a computer-aided diagnosis system.Due to the variable size of the nodules,the low resolution of CT images,the disturbance of noise or other pulmonary tissues,and the high similarity between some benign and malignant nodules,the classification method is difficult to obtain good experiment results.In this paper,the above problems are further studied and the main research contents are as follows:(1)The classification method of benign and malignant pulmonary nodules based on convolutional neural network technology is proposed.The model includes a backbone network,a classification network and a regression network.The backbone network is used to extract the features of nodule images.And it has fewer network layers,which is used to alleviate the problem of small number of nodule samples.Considering the problem of disturbing factors in nodule images,the backbone network specially uses wide-window images and narrow-window images as inputs to extract multi-angle features.In order to reduce the training difficulty of the backbone network and the classification network,a pre-training way is designed.The model increases a regression network which is used to output the diameter of nodule and will be trained with the backbone network in advance to drive the backbone network to focus on the area where the nodule is to extract features.The experiment compares this method with other methods based on traditional features,and the results prove the effectiveness of the neural network features extracted by this method.(2)The classification method of benign and malignant pulmonary nodules that integrates improved Res Net and multiple supervision information is proposed.Themodel includes a backbone network,a decoder,a main classification network and an auxiliary classification network.The backbone network is constructed based on Res Net and then improved to extract the more representative features of nodules.To improve the pre-training way of the previous method,this method combines the backbone network and the decoder as a convolutional autoencoder.According to the diameter of nodule,images which only contain nodules are constructed and used as the expected output of the autoencoder.By training the autoencoder,the backbone network can be adapted to focus on the appearance of the nodule to extract features.The main classification network of this method only uses the malignancy of the nodule as supervision,and the auxiliary classification network uses other diagnosis results of the nodule as extra supervision.The results of the two classification networks are combined to obtain the final results.Experiment results show that this method further improves the classification accuracy compared with the previous method and is effective compared with other recent methods.