FFNET: An End-to-End Framework Based on Feature Pyramid Network And Filter Network for Pulmonary Nodule Detection

Accurate nodule detection with high sensitivity is essential for early lung cancer diagnosis. Focusing on small nodule detection, we propose an end-to-end framework, which includes a backbone, a candidate detection network, and a filter network. The backbone learns multi-layer features so that the region proposal network with feature pyramid structure detects nodules of various sizes, especially small ones. Moreover, the filter net is designed to further classify the proposals with low confidence, which utilizes the decoupled feature maps to make the features of nodules more discriminative. We validate our framework on the LUNA16 dataset. The results show that our framework detects more small nodules, and achieves a comparable performance with other CAD systems.