Steered Convolutional Neurons to Better Learn the Classification of Retinal Vessels

Classification of retinal vessels should not depend on their orientation in the image. Yet, despite characterizing many medical semantic segmentation tasks, equivariance to rotation is not a property of CNNs and is often solely ensured by including random rotations in the data augmentation pipelines. Inspired by non-learning algorithms for vessel classification, we propose to reformulate traditional convolutional neurons using steerable filters to orient their kernels pixel-wise and keep them aligned with the vessels' local orientations. We show that drop-in replacement of standard convolutional neurons with the proposed steered convolutional neurons improves the classification of retinal arterioles and venules in every tested configuration, enhances the classification model's parameter efficiency and strengthens its generalisation abilities in low-data regime.