Subsampling strategies to improve learning-based retina vessel segmentation