Exploiting Residual Edge Information in Deep Fully Convolutional Neural Networks For Retinal Vessel Segmentation

Abstract

Accurate automatic segmentation of the retinal vessels is crucial for early detection and diagnosis of vision-threatening retinal diseases. A new supervised method using a variant of the fully convolutional neural network is proposed with the advantages of reduced hyper-parameters, reduced computational/memory requirements, and robust performance in capturing tiny vessel information. The fully convolutional architectures previously employed for vessel segmentation have multiple tunable hyperparameters and difficulty in end-to-end training due to their decoder structure. We resolve this problem by sharing information from the encoder for upsampling at the decoder stage, resulting in a significantly smaller number of tunable parameters and low computational overhead at the train and test stages. Moreover, the need for pre- and post-processing steps are eradicated. Consequently, the detection accuracy is significantly improved with scores of 0.9620, 0.9623, and 0.9620 on DRIVE, STARE, and CHASE DB1 datasets respectively.