Pulmonary Nodule Detection In CT Images Using Deep Learning Method

Lung cancer,with the lowest survival rate of only about 16% in a five-year span,is the leading cause of cancer-related deaths worldwide.If lung cancer could be diagnosed at an early stage,its 5-year survival rate would be raised to approximately 70%.Thus,early diagnosis is of considerable importance for improving the survival rate of patients.High-resolution lung CT(Computed Tomography)image,which has high precision and is widely used,is an important method for early lung cancer detection and diagnosis.Pulmonary nodule is usually a symptom in early-stage lung cancer so it is highly significant for nodule detection and early treatment with the aid of CT scans.Due to the complexity of inner structures and the similarity of CT values of these varied lung tissues,detection of pulmonary nodules becomes extremely difficult.It is helpful for radiologists to do pulmonary nodule detection and diagnosis combined with computer analysis and image technology.Furthermore,with the help of Computer-Aided Diagnostic(CAD)method,the workload of radiologists could be eased and the misdiagnosis would be reduced,which is the reason of increasingly studies of pulmonary nodules detection based on CT recently.This study mainly focuses on two procedures(pulmonary nodules candidate detection and false positive reduction)of pulmonary nodules detection in lung CT images.Compared with traditional means of detection,the method proposed here does not need pre-designed or preselected nodule features.Instead,it directly uses deep learning model for the detection of candidate pulmonary nodules and the elimination of false positives.The main content of this paper is as follows:(1)An improved Faster R-CNN framework is used in pulmonary nodules candidate detection to improve detection result.According to features of pulmonary nodules and Faster R-CNN framework,three improvements are proposed in this part: adding deconvolution layer after shared part in Faster R-CNN,using ResNet in Faster R-CNN and combining ResNet and Feature Pyramid Networks(FPN)in Faster R-CNN.As proven by experiment,the method combining ResNet and FPN in Faster R-CNN produces the best results in these three improvements because FPN can make full use of semantic information and position information by combining different feature maps in different layers via a top-down pathway and lateral connections.Because of the great capacity of deep learning in absorbing high dimensional features,the conventional procedure of lung parenchyma segmentation can subsequently be removed.(2)A new convolutional neural network is designed to reduce false positives.In this part,input patch size is determined before two networks is designed.Experiments show that VGGlike network is more suitable for false positive reduction.(3)According to the results in the steps of pulmonary nodules candidate detection and false positive reduction,the pulmonary nodule detection procedure is explained in this part.The sensitivity is 89.5% and false positives per scan is 15.2 when using ResNet and FPN in Faster R-CNN in pulmonary nodules detection and VGG-like network in false positive reduction in LIDC dataset.Experiments show the proposed method achieves a good result and omit the procedure of lung parenchyma segmentation and the process of screening pulmonary nodules feature set,so that the complexity of nodule detection can be reduced.