A U-Net Based Progressive GAN for Microscopic Image Augmentation

Dealing with limited medical imagery data by deep neural networks is of a great concern. Obtaining large-scale labelled images requires expertise, is laborious and time consuming, and remains a challenge in medical applications. In this paper, we present a data augmentation method to cope with scarcely available medical imagery data. We propose a U-Net based generative adversarial network to synthesise microscopic images. We adopt a progressive training strategy to guide the synthesising process at multiple resolutions. This also stabilises the training process. The proposed model has been tested on three public datasets and quantitatively evaluated in terms of classification, detection and segmentation performances. Results suggest that training with the proposed augmentation method can provide significant improvements on limited and imbalanced datasets.