Detection of low‐dose computed tomography pulmonary nodules based on 3D CNN‐CapsNet

Abstract

AbstractLung cancer is currently one of the diseases with the highest mortality rate. Early detection of pulmonary nodules is consistently one of the most effective ways to improve the overall survival of cancer patients. However, the existing deep learning‐based pulmonary nodule detection methods still have some problems such as low sensitivity, high false positives, and difficulty in detecting small nodules. To solve the above problems, a low‐dose computed tomography pulmonary nodule detection algorithm based on 3D convolution neural network and capsule network, namely 3D CNN‐CapsNet, is proposed in this work. Combination of full CNN and capsule network reduces the dependence of CNNs on a large amount of data. In 3D CNN‐CapsNet, the convolution kernel of different sizes is applied to the features of different channels to extract richer context information. Then, fused features of different scales are fed into the capsule network for representative feature extraction and more accurate classification. The authors evaluate their proposed method on Early Lung Cancer Program dataset. The nodule detection rate is 95.19%, the sensitivity is 92.31%, the specificity is 98.08%, and the F1‐score is 0.95 which are much better than other baseline methods. These experiments demonstrate that 3D CNN‐CapsNet can effectively improve the detection accuracy of nodules, and can better meet the diagnostic needs of pulmonary nodules.