Space to depth convolution based discriminative pulmonary nodules classification

Lung cancer has become one of the most severe cancers in the world. The detection and classification of pulmonary nodules is of great importance for the lung cancer diagnosis in early stage. However, traditional pulmonary nodule identification methods require doctors to locate the lesion in hundreds of CT images, which is very time-consuming and has the problem of missed diagnosis. In addition, CT images of pulmonary nodule have low resolution and little interclass variability, which affects the performance of the model. Existing deep learning based methods can't solve it well. Therefore, we propose a space to depth convolution neural network (SPD-CNN) based classification algorithm to implement pulmonary nodule classification automatically and accurately. In particular, in the proposed algorithm, we introduce a non-strided module called space to depth convolution (SPD-Conv) to extract and refine feature maps. Moreover, we also utilize convolution based attention module (CBAM) to enable the model to concentrate on the critical features related to pulmonary nodules classification. Simulation results show that the proposed SPD-CNN algorithm can achieve higher classification accuracy than the compared baselines.