Data Augmentation of Thyroid Ultrasound Images Using Generative Adversarial Network

Ultrasound (US) has been investigated as a common method of computer aided diagnosis because of its low-cost, harmless and real-time scanning. Also the rapid development of deep learning segmentation and classification models alleviates the influence of low signal-to-noise ratio and artifacts of ultrasonic imaging. However, due to the privacy issues of medical data, it is not easy to acquire sufficient data for deep learning model training. In recent years, generative adversarial networks (GANs) are widely used in data augmentation. However, GANs suffer from the problem of mode collapse in the training process then generate images with a limited variety. On the other hand, variational auto-encoder (VAE) is free from mode collapse but it generates blurred images. In this work, we study an auto-encoding generative adversarial network combining the advantages of GAN and VAE to generate realistic images for medical thyroid ultrasound image augmentation. Experiment results show that the generated images can simulate realistic ultrasound features and thyroid tissues for augmentation and help training a U-Net model to get better segmentation results.