Pulmonary Nodule Detection Based On YOLOv3 Cascade Neural Network Model

Lung cancer is a disease with high mortality in the world.Although lung nodules can be clearly identified by a CT or X-ray of the lung,there is a huge workload for radiologists and a risk of misdiagnosis.Computer aided diagnosis(CAD)is an effective method to detect pulmonary nodules from CT scan.The development of artificial neural network makes the detection of pulmonary nodules more accurate.In view of the present algorithm based on neural network depends on the large sample,training occupies large memory and take too long,this article is based on cascade CNN-LSTM and YOLO used for screening lung nodules,and false positives to reduce neural network structure,increase the efficiency of data sets,realize the lightweight of neural network,solved the high demand for data sets and computer resources.The cascade neural network model based on YOLOv3 mainly contains two components:1)the YOLO network is used to detect the candidate pulmonary nodules;2)cascaded CNN-LSTM network is used to reduce false positives.Firstly,candidate nodules were detected by YOLO network and their locations were recorded.Then,the 3D images of detected pulmonary nodules were cut into serialized 2D sections.Then,the relevant information between 2D images was extracted by cascading CNN-LSTM network to reduce false positives.In this paper,the idea of image serialization and cascading CNN-LSTM is proposed,and the optimization of neural network is described.The experiment based on the LIDC-IDRI data set shows that the classification accuracy of the model for pulmonary nodules is 91.78%,and the AUC(area under ROC curve)is 93%.During the training,less samples and memory space are used,and the convergence speed is faster.