Class-aware data augmentation by GAN specialisation to improve endoscopic images classification

Abstract

An expert eye is often needed to correctly identify mucosal lesions within endoscopic images. Hence, computer-aided diagnosis systems could decrease the need for highly specialized senior endoscopists and the effect of medical desertification. Moreover, they can significantly impact the latest endoscopic classification challenges such as the Inflammatory Bowel Disease (IBD) gradation. Most of the existing methods are based on deep learning algorithms. However, it is well known that these techniques suffer from the lack of data and/or class imbalance which can be lowered by using augmentation strategies thanks to synthetic generations. Late GAN framework progress made available accurate and production-ready artificial image generation that can be harnessed to extend training sets. It requires, however, to deal with the unsupervised nature of those networks to produce class-aware artificial images. In this article, we present our work to extend two datasets through a class-aware GAN-based augmentation strategy with the help of the state-of-the-art framework StyleGAN2-ADA and fine-tuning. We especially focused our efforts on endoscopic and IBD datasets to improve the classification and gradation of these images.