Dual Path Nets with A New Multi-scale Backbone Architecture For Pulmonary Nodule Detection

摘要

This article mainly completed the detection of lung nodules using deep neural network and built a complete end-to-end detection system. U-net structure has significant advantages in image segmentation. We propose a dual-path network with a U-shaped structure that combines the advantages of residual learning and intensive learning for the detection of lung nodules. And on this basis, the residual module of Resnet was modified. The feature image fed into the residual module is divided into multiple blocks for convolution. Smaller residual blocks are added to each part to broaden the receptive field of each layer. So, it has the ability of extracting multi-scale feature. This network structure is called Res2net. The processed feature images are fused. We also added the SE module to focus on the relationship between feature information channels. It can strengthen the role of useful channels and suppress the output of useless channels. These changes significantly improve the detection effect of lung nodules. We used the position regression of lung nodules to return the predicted position of lung nodules and achieved good results. The detection indicator FROC is significantly better than other detection projects.