Detection And Diagnosis Of Lung Nodules Based On Deep Learning

Nowadays,cancer is a major threat to human life and health.Among them who died of cancer,the proportion of new cases of lung cancer ranks the highest and the mortality rate is the highest.According to this,the research and exploration of lung cancer is of profound significance to all mankind.Lung cancer is a cancer with high incidence and low survival rate.The best current solution to lung cancer remains to detect and treat lung cancer in its early stage and treat it promptly.Early stage of lung cancer is diagnosed through the detection and screening of lung nodules in the analysis of chest computed tomography(CT)images,which has been widely used for lung cancer detection.The manual identification of lung nodules in CT by physicians is time-consuming and laborious,and the results are highly influenced by physician subjectivity.Therefore,many lung nodule detection systems of machine learning have been proposed to improve the automation of lung nodule detection and diagnosis.However,due to the characteristics of pulmonary nodules and the limitations of machine learning,it has been difficult to achieve a major breakthrough.In recent years,the wide application of deep learning in image recognition has proved the superiority of deep learning in image processing.In order to take full advantage of deep learning in image recognition,we propose a deep learning-based method for lung nodule detection and diagnosis in this paper.The main work of this paper includes the following two points.(1)In the paper,we propose a U-Net-based Dense Inverse-residual dual-path lung nodule detection method.In the first stage,we apply the Dense Inverse-residual(DIR)blocks with a fused U-Net structure to learn the potential features of lung nodules.Deeply separable convolution and shortcut connection structures are used in the DIR blocks to enhance the learning of deep latent features by the network while reducing the number of parameters.After the feature-learning network,candidate detection boxes will be generated using a region generation network(RPN),and after filtering them,candidate detection regions will be classified and bounding box regression will be performed.In this paper,experiments were conducted at LUNA16,and the present model is able to identify more lung nodules with a lower false positive rate than other lung nodule detection methods(Res Net50,Deep Lung).In addition,a comparative analysis of the ablation experiments experimentally demonstrats that the internal structure of the method significantly improves the detection of lung nodules.Complexity analysis of the model shows that the method has a smaller number of parameters.(2)A channel attention-based lung nodule classification model,SE-DIRN,is proposed.In the problem of lung nodule classification,in order to exclude the interference of lung tissue on the features learning of lung nodule classification,the article proposes a dual-path network based on the SE module(Squeeze-and-Excitation Block),SE-DIRN.The model efficiently achieves classification of lung nodules by focusing attention on the learning of beneficial features in the feature extraction network through the attention module,which reinforces information about lung nodules and the context.The model is validated on the dataset LUNA16 and compared with other lung nodule classification algorithms.The results show that the method in this paper performs well in classifying lung nodules,achieving 90.2% in accuracy and obtaining an AUC score of 95%.