Multi-discriminator Generative Adversarial Networks for Improved Thin Retinal Vessel Segmentation

Retinal vessel segmentation is an important step in clinical analysis of fundus images. Low contrast and the imbalanced pixel ratios between thick and thin vessels make accurate segmentation of the thin vasculature extremely challenging. In this paper, we present a novel multiscale segmentation method named Multiple discriminator generative adversarial network (MuGAN). MuGAN contains multiple discriminators with different effective receptive fields, which are sensitive to features at different scales. These discriminators jointly teach the segmentation (generator) network to pay attention to multiscale patterns. In addition, multiple discriminators allow our model to incorporate multiple inputs, such as edge enhanced vessel images, during training. We evaluated our method on the publicly available DRIVE and STARE datasets. MuGAN achieved an overall area under the Receiver Operator Characteristic Curve (AUC) of 0.979 for DRIVE and 0.981 for the STARE dataset. On segmenting thin retinal vessels, MuGAN showed quantitative and qualitative improvements on baselines.