Generating synthetic intestine images

Abstract

[en] Capsule endoscopy is a non-invasive medical procedure used to record images of the gastrointestinal tract. While this method is a better alternative for patients, it presents a difficulty to doctors who need to go over as much as 50000 images. Scientists are developing machine learning algorithms that will automatically throw away images free of any anomalies. Like other medical applications, however, available data to train such models is sparse. Therefore, we attempt to create synthetic images that can be used as substitution. For the purpose we have used generative adversarial networks (GANs) as they have recently shown great promise for problems like this one. Training a classifier on both the real and synthetic data, we achieve an increase in the classification accuracy for a dataset of intestine images.