Augmenting data when training a CNN for retinal vessel segmentation: How to warp?

Abstract

The retinal vascular condition is a trustworthy biomarker of several ophthalmologic and cardiovascular diseases, so automatic vessel segmentation is a crucial step to diagnose and monitor these problems. Deep Learning models have recently revolutionized the state-of-the-art in several fields, since they can learn features with multiple levels of abstraction from the data itself. However, these methods can easily fall into overfitting, since a huge number of parameters must be learned. Having bigger datasets may act as regularization and lead to better models. Yet, acquiring and manually annotating images, especially in the medical field, can be a long and costly procedure. Hence, when using regular datasets, people heavily need to apply artificial data augmentation. In this work, we use a fully convolutional neural network capable of reaching the state-of-the-art. Also, we investigate the benefits of augmenting data with new samples created by warping retinal fundus images with nonlinear transformations. Our results hint that may be possible to halve the amount of data, while maintaining the same performance.