Data Augmentation is More Important Than Model Architectures for Retinal Vessel Segmentation

While various deep learning models have recently been proposed or applied to improve the segmentation of vessels in retinal images, the performance gap between different models are often quite small. Such small difference may come from their limited generalization capabilities due to small training data. By simply augmenting data with oriented image patches extracted from the limited training images, we are surprised to observe that even a very simple U-Net with these augmented training patches can outperform the state-of-the-art models with much more complicated architectures or training schemes, and initial gaps between models have become negligible or disappeared. This suggests that it might be more crucial to explore effective data augmentations to extract richer visual information from limited training data, rather than solely focusing on developing other novel deep learning techniques for retinal vessel segmentation.