Attribute-Enhanced Capsule Network for Pulmonary Nodule Classification

Abstract

Purpose

Most pulmonary nodule classification methods solely rely on nodule images without considering the corresponding attribute information. Additionally, conventional convolutional structures often fail to capture the relative spatial relationships among different parts of the lung nodule.

Methods

This paper proposes a pulmonary nodule classification method based on attribute privilege and capsule networks. In this approach, eight attribute characteristics of pulmonary nodules are utilized as privileged information during the identification of benign and malignant cases. This additional information empowers the network to distinguish between benign and malignant aspects of pulmonary nodules more accurately. The capsule structure is introduced to help extract and understand the spatial relationships between different parts of the lung nodule images, the Res2net structure is adopted to extract multi-scale information from lung nodules, and an attention mechanism is incorporated into the backbone network to enhance the efficiency of feature extraction. The proposed classification method was thoroughly evaluated through a series of experiments using the LIDC-IDRI dataset.

Results

The mean identification accuracy of pulmonary nodules reaches 91.6%. This outcome demonstrates that the proposed method is capable of identifying benign and malignant pulmonary nodules with high accuracy.

Conclusion

The novel lung nodule recognition method based on attribute privilege and capsule network contributes to achieving better feature extraction and addressing the challenge of small training datasets for pulmonary nodules.