Research Of Pulmonary Nodule Detection Based On Deep Learning

The incidences of lung cancer still increasing year by year in our country,which brings serious threat to human health.Lung CT diagnosis has been shown to be one of the most effective ways to reduce the incidence of cancer because it can help medical experts to diagnose the disease before it presents any signs or related symptoms.Pulmonary nodule is an early indicator of lung cancer,and the size of nodule is used as an evaluation standard of benign and malignant lung tumors.However,since the early pulmonary nodules are small,it is difficult for even experienced radiologists to find them,which poses a challenge to the early management of lung cancer.In addition,due to the large amounts of lung CT,a large number of experienced doctors are required to screen CT images for a long time,which brings great burden to the work of doctors.Traditional computer-aided diagnosis has achieved remarkable results in nodule detection task,but compared with the method based on deep learning,the detection accuracy still needs to be improved and the detection effect is relatively poor.Based on the existing deep learning methods,an improved pulmonary nodule detection system based on YOLO-v3 and 3D convolutional neural network is proposed in this paper,which not only improves the detection accuracy of pulmonary nodules,but also reduces some unnecessary processes in traditional methods,thus improving the detection speed.The detection process of the system can be divided into two parts: detection of pulmonary nodules and false positive screening.In the stage of pulmonary nodule detection,we improved and optimized the model based on YOLO-v3.Firstly,the spatial attention model is introduced into the residual module of darknet53 to screen the possible pulmonary nodule regions in the image,and the location of pulmonary nodules can be accurately located;Then,a Dense Block is added to the residual module to realize the reuse of image features,generate high-quality image feature representation,and reduce the number of model parameters;Finally,a deconvolution operation is added to enlarge the image resolution to obtain the image feature representation with stronger semantic information to predict nodules.In the stage of pulmonary nodules false positive screening,we propose to use 3D convolutional neural network to capture the global context information of pulmonary nodules,and introduces the self-attention model into the network to refine the features,and uses the idea of skip connect of residual network to fuse the refined feature representation with the original feature to generate the nodule feature representation with stronger semantic information to reduce false positive.The method proposed in this paper is trained and tested on LUNA16 dataset,and the experimental results are analyzed.The experimental results show that the proposed model achieves high accuracy and recall rate.