Detection And Classification Of CT Images Of Pulmonary Nodules Based On Deep Learning

In the modern research field of lung cancer diagnosis,the detection and classification of pulmonary nodules have been the focus of many scholars.Due to the heterogeneity of pulmonary nodules and the visual similarity with the surrounding environment,the feature extraction stage of traditional methods is not only time-consuming and laborious,but also ineffective,which challenges the detection and classification of pulmonary nodules.In recent years,with the rise and wide application of deep learning,researchers in the medical field have paid close attention to it and found a new breakthrough for the early diagnosis of lung cancer,effectively making up for the deficiency of traditional medical technology.Traditional methods for detection and classification of pulmonary nodules are cumbersome,and it is difficult to manually extract features,so that the accuracy is low,and the task of identifying pulmonary nodules in clinical application cannot be well completed.Therefore,a deep learning method is proposed to solve these problems,which is to automatically extract nodule features for learning,detection and classification.Its main research work is as follows:Firstly,in order to reduce the noise interference,lung parenchyma segmentation based on original CT image data was performed in the data preprocessing stage.Secondly,using 3d Faster R-CNN(DPN),which is a combination of Faster R-CNN and dual path network(DPN)to detecte pulmonary nodules,but this process is very time-consuming and cannot be done quickly,it brings a lot of inconvenience to the verification and improvement of the model performance.To more efficiently detecte pulmonary nodules,add the idea of depthwise separable convolution in mobilenet,which greatly improves the detection speed of nodules and makes the detection more efficient,the final total recall rate was 95.36%.Then,deep 3d DPN was used for the classification of pulmonary nodules,in order to improve the classification accuracy,the traditional average pool was replaced by central pool in the classification network.In addition,the idea of multi-scale in inside-outside network(ION)was changed into cross-layer multi-scale feature fusion,so as to extract features better and achieve better classification effect,its accuracy increased from 88.74% to 90.22%.Finally,the results of detection and classification are visualized to further illustrate the effectiveness of the detection network and classification network.In addition,the influence of the initial learning rate,learning rate attenuation strategy and activation function on the experimental results is also studied.