Detection And Classification Of Pulmonary Nodules Based On Convolutional Neural Networks

Lung cancer has the highest incidence and mortality rate of any of the 36 existing cancers,and the death rate from lung cancer has continued to rise worldwide in recent years.Computer Tomography(CT)is a technique that enables high-resolution imaging of lung tissues while presenting rich information on lung nodules,of which malignant lung nodules are the main cause of most lung cancers,so currently lung cancer screening is mainly done by the recognition of lung CT images.With the development of artificial intelligence technology,deep learning-based lung cancer assisted diagnosis systems can quickly and accurately detect and classify lung nodules in CT images to assist doctors in lung cancer diagnosis.In order to further improve the detection efficiency and classification reliability of the lung cancer assisted diagnosis system,this thesis investigates the application of densely connected structures for lung nodule feature extraction based on convolutional neural networks,and constructs a process-oriented lung cancer assisted diagnosis system by integrating detection networks and benign and malignant classification networks.Specifically,this study proposes a topology that can be used for feature extraction,i.e.densely connected microblocks with jump connections.Compared with conventional convolutional neural networks,the proposed densely connected microblocks exhibit better compactness,higher parameter efficiency,and are easier to train.In addition,the use of jump connections for bridging between the microblocks allows the network to efficiently reuse CT image features and achieve effective mitigation of the gradient disappearance problem.In building the lung nodule detection sub-network,in order to make full use of the3 D nature of lung CT data and improve the effectiveness of detection,this study used a3 D Faster R-CNN as the main framework and added 14 improved densely connected micro-blocks to its backbone network part to achieve CT feature extraction.In building the lung nodule classification sub-network,10 densely connected micro-blocks proposed in this study were again used to construct a depth feature extractor to obtain 3D depth features of the nodules themselves.This was then fused with the lung nodule dimensions and the original lung nodule pixels,and finally the benign and malignant dichotomous classification of lung nodules was completed with the help of a gradient boosting tree classifier.To validate the reliability and superiority of the lung cancer assisted diagnosis system proposed in this study,training,testing and evaluation were conducted on a publicly available lung nodule dataset,and multiple sets of comparison experiments were designed.The results show that the maximum sensitivity,average sensitivity and average accuracy of the detection sub-network proposed in this paper can reach 95.2%,86.1%and 79.5% respectively,which is better than the traditional convolutional net-based detection network.In addition,the accuracy,sensitivity and specificity of the classification sub-network proposed in this paper can reach 90.4%,85.86% and 93.37%respectively,which is better than the traditional convolutional network-based classification network.