A Comparative Analysis of Eleven Augmentation Techniques for Enhanced Retinal Pathology Recognition

Deep convolutional neural networks have had considerable success in many computer vision tasks. However, one of the drawbacks of these networks is that they are highly dependent on data volume. Unfortunately, data sparsity affects many interesting application areas. One such area is the recognition of retinal pathologies such as glaucoma and macular degeneration. This study focuses on how to augment the data to overcome the problems of limited data, assume the character of massive data and improve the accuracy of learning models, and finally avoid the problem of overfitting. In particular, we propose to quantitatively study the impact of eleven of the most commonly used augmentation approaches in the particular case of two retinal pathologies. Indeed, these techniques have been developed for generic computer vision problems, but never compared in the particular case of high resolution medical images, as it is the case for the retina. A thorough evaluation of these approaches on two datasets allowed for a rich and extensive discussion and led to several potential perspectives on the subject.