Deep Encoder–Decoder Neural Networks for Retinal Blood Vessels Dense Prediction

Automatic segmentation of retinal blood vessels from fundus images is of great importance in assessing the condition of vascular network in human eyes. The task is primary challenging due to the low contrast of images, the variety of vessels and potential pathology. Previous studies have proposed shallow machine learning based methods to tackle the problem. However, these methods require specific domain knowledge, and the efficiency and robustness of these methods are not satisfactory for medical diagnosis. In recent years, deep learning models have made great progress in various segmentation tasks. In particular, Fully Convolutional Network and U-net have achieved promising results in end-to-end dense prediction tasks. In this study, we propose a novel encoder-decoder architecture based on the vanilla U-net architecture for retinal blood vessels segmentation. The proposed deep learning architecture integrates hybrid dilation convolutions and pixel transposed convolutions in the encoder-decoder model. Such design enables global dense feature extraction and resolves the common “gridding” and “checkerboard” issues in the regular U-net. Furthermore, the proposed network can be efficiently and directly implemented for any semantic segmentation applications. We evaluate the proposed network on two retinal blood vessels data sets. The experimental results show that our proposed model outperforms the baseline U-net model.