GUNet: A GCN-CNN Hybrid Model for Retinal Vessel Segmentation by Learning Graphical Structures

In the retinal vessel segmentation task, maintaining graphical structures of vessels is important for the following analysis steps. However, this task is challenging due to the tiny structures of vessels and bad image quality. Existing methods based on Convolutional Neural Networks (CNNs) can capture local appearances from the regular image grid but are limited to learning high-level features from graphical structures. Motivated by Graph Convolution Networks (GCNs) which capture information on graphs, we propose a novel GCN-CNN hybrid U-shaped model, namely GUNet, which is capable of extracting graphical information of vessels. The hybrid model inherits both merits of CNNs and GCNs. The convolutional blocks extract basic feature representations from local appearances while the GCN blocks learn high-level long-range graphical features along vessels at a deep level. To obtain graphs which effectively represent vessels, we constructed graphs based on the preliminary vessel skeleton segmentation followed by a Hessian filter for vessel enhancement. GUNet takes raw images and corresponding graphs as input and can be trained in an end-to-end manner. The proposed method is evaluated on two fundus photography datasets (STARE and CHASE) and one Scanning Laser Ophthalmoscopy (SLO) dataset (IOSTAR). We conduct experiments to demonstrate that the GCNs module brings significant benefits in terms of graphical similarity and further leads to better overall performances. GUNet also achieves competitive performances compared with state-of-the-art methods.