Research And Application Of Deep Learning Based Lung Nodule Detection Method

Among the cancer diseases found so far,the most dangerous to human life and health is lung cancer.According to the survey,the five-year survival rate of lung cancer patients in China is only about 19.7%,and the incidence and mortality rate of lung cancer is the highest among all cancers.Lung nodules are the main clinical feature of lung cancer in its early stage;therefore,early screening for benign and malignant lung nodules can improve the detection rate of lung cancer and prolong the life span of patients.Traditionally,lung nodule screening is usually performed manually,which is limited by physicians’ professionalism and clinical experience and may result in a certain probability of missing or misdiagnosis.With the application of artificial intelligence in computer-aided diagnosis,it has improved the detection rate of pulmonary nodules while easing the diagnostic work of physicians.In this paper,we investigate the lung nodule detection algorithm based on deep learning,and the main work is as follows:1.An improved Otsu segmentation algorithm is proposed for lung parenchyma extraction.Based on the original Otsu algorithm,the mean gray value and mean variance are introduced to measure the uniform distribution of the gray value of pixels within a class,and the optimal threshold is found to complete the segmentation of lung parenchyma.And compared with the original Otsu algorithm and region growth algorithm,the advantages and disadvantages of various segmentation algorithms were measured using region consistency.The experimental results show that the improved Otsu algorithm extracts higher quality lung nodules in lung CT images.2.A lightweight YOLOv4 lung nodule detection algorithm is proposed.Based on the original YOLOv4 network model,its backbone network CSPDarknet53 is replaced by the lightweight Mobile Net V3 network to reduce the hyperparameters involved in network training and accelerate the convergence of the model.The bilinear interpolation method is used in feature fusion instead of the upsampling process,which reduces the training difficulty of the model.And the semantic information of the top feature map and the location information of the bottom feature map are stacked in a tensor to form a higher channel feature tensor,abandoning the way that the corresponding locations of feature maps in the feature pyramid are superimposed on each other.The network performance is improved by changing the feature map dimension with fewer parameters while reducing the redundancy of parameters.The experimental results show that the improved algorithm in this paper has an AP value of 88.64% and an accuracy of 91.42%.Compared with other target detection algorithms,it has improved in detection accuracy and AP.3.An intelligent lung nodule detection system was designed and implemented.From the perspective of system design,the functional requirements of the system were analyzed,the lung nodule detection methods used in this paper were integrated,and a lung nodule detection system was implemented and tested for relevant functions.The test results show that the system can quickly complete the task of detecting pulmonary nodules and assist doctors in reaching diagnostic results.