Research On Detection And Diagnosis Algorithm Of Pulmonary Nodule Based On Convolutional Neural Network

Lung cancer is currently one of the diseases with the highest morbidity and mortality rates all over the world,which is a serious threat to human life.Early diagnosis and treatment of lung cancer is the key to improve the five-year survival rate of lung cancer patients.Pulmonary nodules are the main manifestation of lung cancer in the early stage.Detection,segmentation,and benign and malignant classification of pulmonary nodules based on CT imaging are the key steps and effective methods for early diagnosis of lung cancer.Studying the pulmonary nodule diagnosis algorithm based on convolutional neural network and developing the corresponding system can provide theoretical basis and means for the early diagnosis of lung cancer.The detection and diagnosis of pulmonary nodules usually include three steps.1.pulmonary nodules detecting;2.pulmonary nodules segmentation based on the results of the detection;3.benign and malignant classification of pulmonary nodules based on the location information and mask information of the nodules.Due to the large differences in the size and morphology of pulmonary nodules in CT images,it is prone to misdetection and missed detection.Therefore,we designed a method based on Two-Stage Convolutional Neural Network(TSCNN)to detect suspicious pulmonary nodules and remove false positive nodules.The convolutional neural network architecture used in the first stage is a UNet network based on the Res Dense structure.In order to improve the recall rate without introducing too many false positives,a segmentation target-oriented sampling strategy is designed and the model is trained using offline hard sample mining methods.On the trained model,a two-scale prediction method is used to quickly locate suspicious nodules.The network architecture in the second stage are three 3D-CNN classification networks based on the designed dual pooling structure,namely Se Res Net,Dense Net,and Inception Net.The average value of their output results is taken as the probability of whether it is a nodule.To alleviate the problem of a small amount of training data and imbalance of positive and negative samples in this classification task,a data augmentation method based on the random mask is proposed.The proposed nodule detection method is experimentally verified on the LUNA16 dataset.Experimental results show that the method has a Competition Performance Metric(CPM)of 92.5%.In order to effectively segment nodules of different sizes,different intensities and different morphologies,a pulmonary nodule segmentation model called a Dual-Branch Residual Network(DB-Res Net)is proposed in this paper,which can obtain the features of pulmonary nodules on different slices and on different scales simultaneously.In view of the large intensity difference in pulmonary nodules,a method of central intensity pooling is proposed to obtain the intensity features,which is stitched with the features extracted using central pooling to form new pulmonary nodule features.In addition,a weighted sampling strategy based on the number of nodule boundary points is also proposed.More sampling of nodules with small size and irregular borders is performed to achieve intensive training of these difficult to segment nodules.The proposed nodule segmentation method is experimentally verified on the public dataset named LIDC-IDRI.The experimental results show that the dice similarity coefficient,average surface distance,sensitivity,and positive prediction value of the method achieved 82.74%,0.19 mm,89.35%,and 79.64%,respectively.Aiming at the problem that clinical classification of benign and malignant pulmonary nodules is difficult,this paper proposes a Multi-Model Multi-Scale Ensemble Learning method based on 3D Convolutional Neural Network(MMMSEL-3DCNN).The proposed method can select multiple models for benign and malignant classification according to the size of nodules.Each model sends the nodule mask region image,nodule ROI image,and its corresponding enhanced image as the three channels of network input to the network for training to extract the advanced features of the nodule.In addition,a data augmentation method called stochastic evolution is proposed for the problem that there is only a little labeled training data in the dataset.This method simulates the natural change process of diseased tissue deterioration and healing,and is implemented according to the local distortion algorithm in graphic imaging,which effectively increases the training data of labeled pulmonary nodules and further improves the robustness of benign and malignant classification models of pulmonary nodules.The experimental verification on the LIDC-IDRI dataset shows that the accuracy and Area Under Curve(AUC)of the proposed method reach 0.905 and 0.942,respectively.Comparative experiments with several methods with the best performance show that the proposed methods for pulmonary nodule detection,segmentation and benign and malignant classification can effectively improve the accuracy of pulmonary nodule diagnosis,reduce the rate of misdiagnosis and missed diagnosis,and provide a theoretical basis and effective means for early diagnosis of lung cancer.