Pulmonary Nodules Detection via 3D Multi-scale Dual Path Network

It has become an epidemic trend to use computer-aided detection (CADe) systems based on deep learning to help physicians improve the efficiency and diagnostic rate of pulmonary nodules. Considering the problems of low sensitivity, high false positives, and difficulty in detecting small nodules in the existing CADe system, this paper proposes a pulmonary nodule detection model based on a 3D multi-scale dual-path network. The proposed model uses a 3D convolutional neural network (CNN) for spatial information extraction to generate more representative features and extracts multi-scale feature information by constructing an Unet-like network structure and adding pre-processing blocks. The model uses multi-scale architecture and dual-path network (DPN) to extract deep features, combines the characteristics of ResNet and DenseNet to improve feature extraction capabilities. In addition, the focal loss is introduced as the classification loss function of the algorithm to alleviate the imbalance problem of positive and negative samples. The results based on the LUNA16 dataset show that this model can improve the diagnosis rate of pulmonary nodules with a low false-positive rate.