Research Of Pulmonary Nodule Detection And Diagnosis Based On Deep Learning

Pulmonary nodule is an early manifestation of lung cancer.The detection and diagnosis of early pulmonary nodules has great clinical significance.It can greatly reduce the incidence and mortality of lung cancer.With the rapid development of CT technology,CT images have become the most effective data form for diagnosing pulmonary nodules.In clinical practice,radiologists diagnose pulmonary nodules mainly through manual reading of CT images.A large number of medical resources are used to handle repetitive tasks.Therefore,computer-aided diagnosis techniques(CAD)have received widespread attention in recent years.It can reduce medical costs,improve diagnostic efficiency,and promote the development of medical services.The early pulmonary nodule intelligence analysis systems mainly based on traditional machine learning algorithms.This type of method is cumbersome and uses artificial design features that restrict the performance of the algorithm.In recent years,with the extensive use of deep learning algorithms,more and more research work has used deep learning algorithms to solve medical problems.The existing intelligent analysis system based on deep learning mainly uses 2D CNN(2D CNN).This method does not make full use of 3D spatial information.It not only restricts the accuracy of the algorithm itself,but also makes the ratio of false positives high..To solve the problems existing in the existing methods,this paper uses 3D CNN as the basic network to extract high-level semantic 3D features,and then design an effective loss function according to specific tasks to optimize the network.The main works of this paper are as follows: 1.3D RPN based on 2D image detection framework Faster RCNN is used to extract the region of interest.Because there are only two categories in this problem,the result generated by RPN already contains Category information.The network structure realizes feature fusion.When the model is optimized,the online difficult sample mining is realized,so that the network can detect difficult samples.2.The 3D CNN and focal loss function are usedto train the false positive suppression model which can remove the non-nodular samples from the region of interest extracted by the 3D RPN and solve the problem of high false positive rate in the traditional CAD system.3.The 3D full convolutional network(3D FCN)is established to segment the lung nodule lesion area in a pixel-to-pixel manner and adversarial learning strategy is utilized to optimize the model.4.One 3D CNN classification network is performed to identify the attributes of pulmonary nodule,such as malignancy,lobulation and spiculation,which can give doctors advices for the final diagnosis.In this paper,3D CNNs are used to automatically extract 3D high-level semantic features and appropriate loss functions are designed to optimize the model.This article mainly consist of three parts,such as pulmonary nodule detection,pulmonary nodule segmentation and attribute classification of pulmonary,which achieves the location,quantitative and qualitative functions respectively to better assist the diagnosis.Finally,a 98.6% lung nodule detection rate was achieved on the LIDC data set.According to the evaluation criteria of the professional lung nodule detection competition LUNA2016,the FROC Score reached 0.957;the Dice score of pulmonary nodule segmentation accuracy reached 0.872;the accuracy of malignant identification reached 95.4%,the accuracy of lobulation reached 91.6%,and the accuracy of spiculation reached 89.6%.