Scale-space approximated convolutional neural networks for retinal vessel segmentation.

•Conventional CNN structures may cause aliasing in multi-scale feature extraction.•In the scale-space approximated CNN (SSANet) decimation is replaced with upsampling.•Upsampling after downsampling is essentially Gaussian blurring in scalespace theory.•Reduced receptive field is offset with residual blocks that also increase capacity.•Proposed SSANet shows state-of-the-art accuracy for retinal vessel segmentation.