Multi-scale Pulmonary Nodule Detection by Fusion of Cascade R-CNN and FPN

Abstract

Accurate pulmonary nodule detection in chest computed tomography (CT) is a crucial step in diagnosis and treatment of lung cancers. However, pulmonary nodules in CT images vary in size and are difficult to detect. In this paper, we proposed a new model fusing Cascade R-CNN and feature pyramid network (FPN) for multi-scale pulmonary nodule. In our model, ResNet is used as the backbone network for feature extraction. Meanwhile, to enhance the semantic features of each layer, the feature pyramid network is constructed by fusing the upper layer features with the lower layer features. Cascade R-CNN is introduced into our model for detection using different Intersection over Union (IoU) thresholds. To improve the average detection precision of small nodules, the Mosaic data are employed to augment the training set. Our model achieved 0.879 average precision (AP) on Lung Nodule Analysis 2016 (LUNA16) dataset, which shows that our model can effectively detect multi-scale pulmonary nodules, and the average precision on nodules of small sizes has increased significantly.