Pulmonary Nodule Classification of CT Images with Attribute Self-guided Graph Convolutional V-Shape Networks

Accurate identification and early diagnosis of malignant pulmonary nodules are critical to improving the survival rate of lung cancer patients. Recently, deep learning methods have been proved to be successful in computer-aided diagnosis tasks. However, most advanced research work does not fully utilized valuable attribute prior knowledge for semantic reasoning to guide the network. Therefore, it lacks interpretability and hence is difficult for clinical radiologists to understand and apply. To comprehensively tackle these challenges, we propose a novel Attribute Self-guided Graph Convolutional V-shape Networks (AS-GCVN) for pulmonary nodules classification with steps as follows. We first develop a sub-network for representation learning, which can effectively extract image-level features. Second, we construct a graph convolution V-shape network to model the semantic information of attributes to guide the classification of benign and malignant pulmonary nodules accurately. Moreover, an Attribute Self-guided Feature Enhancement (ASFE) module is proposed to improve the ability of graph semantic reasoning, which can map the image features extracted by the convolutional neural network to attribute features through adaptive learning. Finally, the two sub-networks effectively integrate attribute inference knowledge and representation learning to enable end-to-end training. This way can further improve the interpretability and robustness of pulmonary nodule classification. Extensive experimental results on the LIDC-IDRI dataset demonstrate that our approach obviously outperforms other existing state-of-the-art methods.