Attentive and ensemble 3D dual path networks for pulmonary nodules classification

Abstract

Automated pulmonary nodules classification aims at predicting whether a candidate nodule is benign or malignant. It is of great significance for computer-aided diagnosis of lung cancer. Despite the substantial progress achieved by existing methods, several challenges remain, including the lack of fine-grained representations, the interpretability of the reasoning procedure, and the trade-off between true-positive rate and false-positive rate. To tackle these challenges, in this work, we present a novel pulmonary nodule classification framework via attentive and ensemble 3D Dual Path Networks. Specially, we first devise a contextual attention mechanism to model the contextual correlations among adjacent locations, which improves the representativeness of deep features. Second, we employ a spatial attention mechanism to automatically locate the regions essential for nodule classification. Finally, we employ an ensemble of several models to improve the prediction robustness. Extensive experiments are conducted on the LIDC-IDRI database. Results demonstrate the effectiveness of the proposed techniques and the superiority of our model over previous state-of-the-art.