Taking Study On Detection Method Of Pulmonary Nodules Based On Deep Learning

At present,among the malignant tumors in the world,the morbidity and mortality of lung cancer are high.COVID-19,which broke out in December 2019,has spread all over the world,and people will cause different types of lung disease after infection.Lung disease can cause many complications,such as tracheitis,heart disease and so on,which seriously endanger people’s health.Therefore,early screening of lung diseases is particularly important.The early clinical manifestation of lung cancer is lung nodules,and computed tomography can be used to detect lung nodules early,so as to achieve the purpose of early detection and early treatment.Considering that lung nodules manifest themselves in the body in a variety of ways,it is very easy to misdetect and miss tests by relying on the doctor’s personal knowledge reserve.In recent years,deep learning has achieved great success in related applications of image processing.This thesis is based on deep learning methods,and the main content includes two parts:lung parenchymal segmentation and lung nodule detection.The specific work is as follows:(1)In view of the current problem of lung parenchymal segmentation when the edge of the lung structure is incomplete when segmented adhesions,this thesis proposes an improved segmentation algorithm based on U-Net structure,which improves the generalization ability and robustness of the network model by adding expansion convolution,residual structure and improving loss function in the coding part.The improved network optimizes the segmentation effect of the lung parenchyma,and can accurately find the specific location of the lung region,including the exact target boundary.Experiments are carried out on the LUNA16 data set,and the final results show that the Dice similarity coefficient of this method is 92.67%,the performance is better than other segmentation methods involved in the experiment,and the segmentation effect is good,which provides a good data basis for the next step of lung nodule detection.(2)in view of the complex background of lung images and the characteristics of small targets of pulmonary nodules,an improved model based on YOLOv5 is proposed to detect pulmonary nodules.The enhanced feature extraction layer of YOLOv5 combines the bi-directional feature pyramid network Bi-FPN,and the attention mechanism module is introduced into the algorithm of Neck.Finally,it can enhance the proportion of target features and reduce the proportion of background features,and effectively make up for the loss of target position in the feature map,so as to improve the detection ability of small targets.Finally,the horizontal comparison of several other classical detection algorithms and ablation experiments show that the average detection accuracy of the optimized detection algorithm proposed in this paper is 91.4%,which is generally improved by more than 6%.In this thesis,the accuracy and speed of lung parenchymal segmentation and pulmonary nodule detection were effectively improved by improving the lung parenchymal segmentation algorithm and lung nodule detection algorithm.