Towards the Substitution of Real with Artificially Generated Endoscopic Images for CNN Training

Abstract

The generalization performance in deep learning is linked to the size and the variations of the samples available during training. This is apparent in the domain of computer-aided gastrointestinal tract abnormality detection, where the lesions can vary a lot from each other and the number of available samples is limited, mainly due to personal data protection legislations. In this work we present a novel approach of tackling the problem of limited training data availability by making use of artificially generated images. More specifically we trained a Generative Adversarial Network (GAN) using Wireless Capsule Endoscopy (WCE) images to generate fake but realistic images from the small bowel. The generated images were then used to train a Convolutional Neural Network (CNN) to identify inflammatory conditions on real WCE images. To evaluate the performance of our approach, in our experiments we compare the generalization performance of the same CNN architecture trained separately with real and fake images, obtaining 90.9% and 79.1% Area Under Receiver Operating Characteristic (AUC), respectively. The results show that training using solely artificially generated data can be effective in cases where real training data are inaccessible.