Texture-embedded Generative Adversarial Nets for the synthesis of 3D pulmonary nodules computed tomography images

Abstract

Lung cancer is the second most common cancer and the leading cause of cancer-related death worldwide. Screening with low-dose computed tomography is crucial to detect early-stage lung cancer. Computer-aided diagnosis (CAD) can help clinicians to make diagnosis more quickly and more accurately. CAD based on deep learning algorithms is gaining attention. These algorithms rely on large amount of training data, which are barely available in the field of medical imaging, therefore data augmentation becomes essential. Generative Adversarial Nets (GAN) is an emerging solution for data augmentation and has been successfully used to generate realistic pulmonary nodules. In this study, we developed Texture-embedded GAN, which took the texture of nodule into consideration by introducing a loss function based on Gabor filters. We trained Texture-embedded GAN with images of 1075 nodule from the LIDC-IDRI dataset. Visual Turing Test showed that Texture-embedded GAN could generate images realistic enough to deceive expert radiologists. Data augmentation with Texture-embedded GAN improved the performance of ResNet-based classifier, which could distinguish benign and malignant nodules with 0.883 accuracy and 0.950 AUC. It was concluded that Texture-embedded GAN could generate realistic pulmonary nodules with sufficient diversity and was useful for data augmentation.