Research On Pulmonary Nodule Detection And Combined False Positive Removal Based On Deep Learning

As the malignant tumor with the highest morbidity and mortality in China,lung cancer poses serious threats and obstacles to human health and social development.Improving the early diagnosis rate of lung cancer is essential for reducing its mortality.Pulmonary nodules are an early manifestation of lung cancer.The pulmonary nodules are the tissues of the lungs.They show white shadows on medical images.The detection of pulmonary nodules is the key to the diagnosis of lung cancer.Chest computed tomography(CT)is a common imaging tool for pulmonary nodule detection and an effective way to diagnose early lung cancer.Because the contrast of lung nodules in CT and X-ray images is relatively low,and there is occlusion overlap with the lungs and other surrounding tissues,and the size and opacity of the nodules themselves,to the pulmonary nodules Detection poses a big challenge.At present,the use of deep learning in the auxiliary diagnosis of medical imaging has been widely studied and applied.For the detection and research of pulmonary nodules,the problems of missed detection and misdetection are more prominent,so the research of detection of pulmonary nodules based on deep learning and remove false positive and improves the quasi-rate of detection,has important significance and practical application value.Based on the subset of the largest common pulmonary nodule dataset LIDC-IDRI,the LUNA16 dataset,the following studies were conducted on the detection of lung nodules and the reduction of false positives:(1)Pulmonary nodule detectionAiming at the problem of pulmonary nodule detection,a detection of pulmonary nodules algorithm based on improved 3D Faster R-CNN was constructed to fully exploit the effective information in the three-dimensional space of the data.To better extract features at multiple scales and increase the number of anchor points,design five different sizes of anchors.In light of the diminutive targets and their great differences in size in the pulmonary nodules detection.The SKNet module is introduced,which combines the features extracted by different convolution kernels in a nonlinear way to achieve the receptive field size.This convenient network adaptive learning feature parameters,adaptive receptive field size adjustment can enhance the network model’s ability to learn related features,and improve the robustness of the network model.(2)Candidate nodule false positive removalAiming at the problem that the candidate nodule false positives generated by the detection of pulmonary nodules network is higher,a joint false positive is proposed based on 3D U-Net network and 3D DCNN network.Introducing dilated convolution replace partial convolution operation in a cavity convolution.While increasing the receptive field,the size of the feature map is kept constant,and the loss of some important information is avoided as much as possible.The high-performance performance of the true and false nodule classification task is completed,and the candidate knot is solved to some extent.The problem of high false positives achieves high-precision lung nodule detection.This paper hopes that through the research of detection of pulmonary nodule sand false positive removal algorithm,it can promote the research of pulmonary nodule detection,more accurately detect the existence of pulmonary nodules,and provide more effective medical assistance for the treatment of patients with pulmonary nodules.