Boundary-enhanced cascaded network for joint optic disc and optic cup semantic segmentation

Clinically, calculating the Cup-to-Disc Ratio (CDR) from fundus images is a straightforward method to screen and diagnose glaucoma. However, the accurate calculation of CDR heavily depends on precise segmentation of the Optic Disc (OD) and Optic Cup (OC). Most U-net-based OD and OC segmentation methods have achieved good results, but these approaches may lead to outliers, poor integrity, inaccurate and unsmooth contours. To alleviate these issues, we propose a Boundary-Enhanced Cascaded Deep Convolutional Neural Network (BEC-DCNN) for robust and accurate OD and OC segmentation in fundus images, which is followed by a two-stage framework. The first stage employs the DeepLabv3+ network with attention mechanism to detect the OD. In the second stage, we integrate a boundary auxiliary branch, which can obtain multi-scale contour feature maps from the backbone network to produce smoother OD and OC segmentation results. This paper designs a simple and efficient joint loss in training to constrain both the segmentation branch and the boundary branch, which is helpful to improve the segmentation precision and the smoothness of the boundary. Our experiments conducted on the three public datasets, Rim-One-r3, Drishti-GS, and REFUGE, demonstrate that our proposed BEC-DCNN outperforms many previous algorithms. Through qualitative and quantitative comparisons with other methods, our proposal produces more accurate and smoother contours. Our method is also applicable to other medical image segmentation tasks that require accurate edges.