Vessel Curvature-Based Data Augmentation Technique for Retinal Fundus Images

A novel vessel structure-based dataset augmentation approach has been proposed in this work to aid the learning of a vanilla U-Net-based model for retinal vessel segmentation. A functional formulation is proposed in this work to represent a fundus image as a mix of background surface and vessel surface. We have individually proposed a separate model for operating the background region and vessel curves independently and validated the augmentation strength of the proposed work by testing the performance of the learned U-Net model on two benchmark datasets, namely, DRIVE and STARE. Finally, we have compared and reported the performance of the U-Net model, trained on the augmented data, based on the metrics of accuracy, specificity, and sensitivity in this work.