A Novel Data Augmentation Method Using Style-Based GAN for Robust Pulmonary Nodule Segmentation

摘要

Accurate and automatic segmentation of pulmonary nodules from computed tomography (CT) images is an important task for lung cancer analysis. However, the scarcity and imbalance of labeled data make it difficult for robust nodule segmentation. In this paper, a novel data augmentation method using style-based generative adversarial network (GAN) is proposed to synthesize augmented training data. Styles and semantic labels are firstly extracted from the whole dataset, then augmented CT images are synthesized for each semantic label with randomly chosen styles. Using a region-aware network architecture, the style in specific semantic regions can be more finely controlled, including lung nodules, vascular structures and pleural surfaces. In our experiments on LIDC-IDRI datasets, we show that the proprosed method can generate nodule samples realistically, and lead to more accurate and robust nodule segmentation with the augmented samples compared to the previous works.