Glaucoma Identification Using Convolutional Neural Networks Ensemble for Optic Disc and Cup Segmentation

Abstract

Glaucoma is one of the diseases that can cause incurable blindness. The first noticeable symptoms appear only when the disease has progressed but an early diagnosis of the disease prevents the severe consequences of disease progression. In this paper, we developed a Convolutional Neural Networks (CNNs) based ensemble for joint optic disc (OD) and optic cup (OC) segmentation using modified Attention U-Net architecture with pre-trained ResNet34, ResNet50, MobileNet, Inceptionv3, DenseNet121 as backbones. The ensemble was trained on a mixed dataset consisting of REFUGE, Drishti-GS, and RIM-ONE r3 (RIM-ONE) datasets of eye fundus images and tested on images of each dataset separately. The most accurate joint OD and OC segmentation is achieved using an ensemble consisting of modified Attention U-Net with pre-trained ResNet34, Inceptionv3, and DenseNet121 as backbones and majority voting for final prediction. The highest Dice of 0.961 for OD and 0.894 for OC is achieved on the REFUGE test dataset, 0.974 for OD and 0.916 for OC on the Drishti-GS test dataset, and 0.978 for OD and 0.902 for OC on the RIM-ONE test dataset. The highest Intersection over Union of 0.925 for OD and 0.808 for OC is achieved on the REFUGE test dataset, 0.950 for OD and 0.845 for OC on the Drishti-GS test dataset, and 0.957 for OD and 0.822 for OC on the RIM-ONE test dataset. Using the segmentation results, the cup-to-disc ratio (CDR) has been calculated to classify eye fundus images into mild-stage, moderate-stage, severe-stage glaucoma, and non-glaucoma cases.

Keywords

• Ensemble networks
• glaucoma
• machine learning (ML)
• segmentation
• voting methods