Pulmonary Nodule Detection Based on Faster R-CNN with Adaptive Anchor Box

Abstract

Early pulmonary nodule detection is very important in lung cancer diagnosis and screening. Most state-of-the-art lung nodule detection models are based on Faster Region-based Convolutional Neural Network (Faster R-CNN) due to its superior performance. However, this object detection approach faces difficulties with the variety of nodule sizes in training datasets. In this paper, we propose a novel Computer-Aided Detection (CAD) system based on Faster R-CNN model with adaptive anchor box for lung nodule detection. Our method employs ground-truth nodule sizes in the training dataset to generate adaptive anchor box sizes of Faster R-CNN. Learned anchors are used as hyper-parameter to boost Faster R-CNN's detection performance. A residual convolutional neural network is proposed to reduce false positives from Faster R-CNN's output. Our method is trained and tested on the largest publicly available LUNA16 dataset. Experiments show that our proposed system achieves a high sensitivity of 95.64% at 1.72 false positives per scan, and a Competition Performance Metric (CPM) score of 88.2%, which outperforms other recent state-of-the-art detection methods. The false positive reduction network achieves a sensitivity of 93.8%, specificity of 97.6% and accuracy of 95.7%. An additional evaluation on a completely independent SPIE-AAPM dataset demonstrates the generalization of our proposed model with 89.3% sensitivity.