Deep Convolutional Nets for Pulmonary Nodule Detection and Classification

In this study, a novel pulmonary nodule detection and classification system with 2D convolutional neural networks is proposed. The objective is to effectively address the challenges in lung cancer diagnosis and early treatment. The system consists of two stages: nodule detection and false positive reduction. For nodule detection, we introduce a detection framework based on Faster R-CNN, which integrates a deconvolution layer to enlarge the feature map and two region proposal networks to concatenate the useful information from the lower layer. In order to ensure high sensitivity, the conditions at this stage are simple and loose. Therefore, a boosting architecture based on 2D CNNs is designed for false positive reduction. In order to improve classification accuracy, every training model pays attention to those data that are not easy to classify. In experiments, our method is conducted on LUNA16 challenge. The sensitivity of nodule candidate detection achieves 86.42%. For false positive reduction, sensitivities of 73.4% and 74.4% at 1/8 and 1/4 false positives per scan are obtained, respectively. It proves that our method can maintain a satisfactory sensitivity even with extremely low false positive rates.