Automatic Detection Of Lung Nodules Based On 3D Convolutional Neural Networks

Lung cancer is the leading cause of cancer deaths worldwide.Its 5-year survival rate can be improved by the early diagnosis and treatment.The detection of lung nodules,the potential precursors to lung cancer,is an essential step in the early diagnosis of lung cancer,thus it is of crucial significance.In this paper,three-dimensional convolutional neural networks are proposed for the lung nodule detection.The networks can automatically learn some high-level features without the need of manually designing features and can achieve high detection sensitivity with a relative low number of false positives.The lung nodule candidates are obtained based on three-dimensional fully convolutional networks.Two networks were developed in the paper: a multi-scale fully convolutional network(FCNet_RF17-33)and U-Net(U-Net_RF33).The networks can learn features at different levels.During the training process,image patches are randomly and equally sampled from nodule regions and other non-nodule regions in the preprocessed CT images and are used as the inputs to the networks.The outputs of the networks are lung nodule probability maps corresponding to the central area of the input image patches.The voxel value in the probability map represents the probability of it being a nodule.In the testing stage,a probability map corresponding to the entire CT image is obtained by a sliding window strategy.Then,the candidate positions of the lung nodules are obtained by post processing the probability map.The dataset released by the LUNA16 challenge is used as the test set.The FCNet_RF17-33 achieves a sensitivity of 97.5% with 68.7 FPs/scan.The sensitivity of the U-Net_RF33 is 96.4% at 14.8 FPs/scan.For the false positive reduction part,two three-dimensional convolutional neural networks are used.The two networks are named Basic8 and Deep11 and are designed to classify the lung nodule candidates.Basic8 has 8 layers including 4 convolutional layers while Deep11 is deeper with 11 layers including 8 convolutional layers.By comparing different combinations of networks,U-Net_RF33 is finally used in the candidate detection stage and Basic8 is adopted in the false positive reduction stage.This is used as the final solution of the proposed computer-aided pulmonary nodule detection system.The structures of the networks are simple.The sensitivity is 94.4% at 2FPs/scan and the competition performance metric is 0.898.