The Identification Of Pulmonary Nodules Based On Deep Neural Network

Lung cancer has the highest incidence and mortality rate among all cancers in the world,and it poses a great threat to human life.Studies have shown that early detection and treatment of lung cancer can reduce the mortality of patients with lung cancer.However,there is no obvious clinical feature of lung cancer in the early stage.It is difficult to find the lung cancer and it is easy to misdiagnose and miss diagnosis.Low-dose chest CT scan is a common method for early screening of lung cancer,but this requires radiologists to carry out manual reading to draw a conclusion,which undoubtedly increases the workload of doctors.Because of manual identification,if the doctor’s clinical experience is not enough,it is inevitable that there will be misdiagnosis or missed diagnosis.With the rapid development of computer and neural network technology in recent years,the auxiliary diagnosis and treatment through medical imaging and machine learning will inevitably occur,the accurate recognition of early lung cancer has become one of the current research hotspots and important directions.In this paper,the recognition technology of pulmonary nodules based on deep neural network is studied,and this method is used to recognize the pulmonary nodules of CT images by radiologists.The research in this paper is divided into three stages: the first stage is to find all suspected pulmonary nodules and segment them;the second stage is the screening of the false positive lung nodules;the third stage is the formation of the lung nodule identification network.In this paper,the method of the convolution neural network is used as the core to study for the identification of the lung nodules.The main contributions are as follows:1.For the segmentation of suspected pulmonary nodules,based on the study of many deep learning methods,an improved U-Net method is proposed in this paper.On the basis of U-Net structure,residual block and dense block are combined to make the networkextract more features of pulmonary nodules.It also have to solve the issue of gradient disappearance,which is helpful to improve the segmentation accuracy of pulmonary nodules.Focal loss cost function is used instead of the traditional cost function,which can solve the problem that the segmentation effect of small target is not ideal,and further improve the segmentation accuracy.2.For the detection of false positive pulmonary nodules,based on the study of many false positive detection algorithms,a multi-scale false positive screening method based on3D-CNN is proposed in this paper.This method improves the existing 3D-CNN model,draws lessons from the idea of VGG,adopts the small convolution kernel to sample,and deepens the network structure and integrates the residual module to improve the performance of the model.In this paper,according to the diameter of pulmonary nodules,the size of input pulmonary nodules in 3D-CNN model is divided more reasonably,and the weight distribution is optimized in the multi-scale model.The experimental results show that the proposed method is superior to the general 2D-CNN and 3D-CNN methods.3.For the formation of pulmonary nodules recognition network,on the basis of the first two studies,the improved U-Net model and 3D-CNN multi-scale pulmonary nodules false positive detection model are connected in series.Combined with pulmonary nodules centroids acquisition algorithm and pulmonary nodular mass acquisition algorithm,the final lung node recognition network is obtained.By comparing with other lung nodules recognition methods,the recognition method in this paper is accurate and sensitive.In terms of specificity,it is superior to the general method.In this paper,CT images of lung can be identified by deep neural network,and pulmonary nodules can be found in the early stage,so that more patients with lung cancer can find malignant pulmonary nodules in early or routine physical examination,and obtain the opportunity of operation.