Automatic Pulmonary Nodule Detection Using Faster R-CNN Based on Densely Connected Network

Abstract

Accurate detection of pulmonary nodules in CT images is a key task in performing computer-aided diagnosis of pulmonary diseases. In this work, inspired by the successful application of Faster R-CNN in object detection and the superiority of dense convolutional networks in feature propagation, we proposed a modified Faster R-CNN with an improved densely connected network as the backbone for lung nodule detection in medical images. In the proposed network, the backbone for feature extraction can be considered as a combination of multiple densely connected micro-blocks with skip connections. Skip connections in the micro-blocks enhances the propagation of features between layers, thus enable feature reusage. These micro-blocks effectively mitigate the problem of gradient vanishing in feature propagation due to their dense properties. In addition, the compact structure of these micro-blocks facilitates the network to extract and learn CT image features more efficiently. Finally, these micro-blocks have fewer parameters and higher parameter efficiency. The proposed method was tested and evaluated on the public lung nodule dataset LUNA16. When a ten-fold cross validation was performed, the proposed network achieved a FROC score of up to 0.952 and a CPM score of up to 0.861. Experimental results show that the proposed network is capable of detecting pulmonary nodules with higher sensitivity and accuracy than other conventional lung nodule detection methods.