Joint Dice and Intersection over Union Losses for Deep Optical Disc Segmentation

Abstract

Optic disc segmentation on retinal images is essential for the diagnosis of various eye-related diseases, such as diabetic retinopathy, glaucoma, and macular edema. However, manual segmentation of optic discs from fundus images by ophthalmologists is time-consuming, tedious, and labor-extensive. In the literature, various optic disc segmentation algorithms have been proposed. In general, the existing methods are evaluated in terms of the Dice coefficient and Area of Overlap. These two metrics indicates the capability of the existing methods in preserving the accurate optic disc contour and size. However, most available segmentation methods do not attempt to reduce the two measures directly. In this paper, we present a new deep optical disc framework that can tackle the drawbacks of the methods in the literature. The proposed framework performs segmentation by taking the advantages of the U-shaped convolutional neural network (U-Net); U-Net could be trained using a limited number of data. Unlike most other methods in the literature, we use a joint dice and intersection over union losses for training the deep network. We train and evaluate the proposed framework on retinal images from the DRIVE and DRISHTI-GS datasets. In the experimental parts, the proposed framework is capable of outperforming competing methods in terms of the Dice coefficient and Area of Overlap. Therefore, our framework is suitable for broad applications of automated retinal diseases diagnosis.