Robust optic disc and cup segmentation with deep learning for glaucoma detection

Abstract

Glaucoma is rated as the leading cause of irreversible vision loss worldwide. Early detection of glaucoma is important for providing timely treatment and minimizing the vision loss. In this paper, we developed a robust segmentation method for optic disc and cup segmentation using a modified U-Net architecture, which combines the widely adopted pre-trained ResNet-34 model as encoding layers with classical U-Net decoding layers. The model was trained on the newly available RIGA dataset, and achieved an average dice value of 97.31% for disc segmentation and 87.61% for cup segmentation, comparable to that of the experts’ performance for optic disc/cup segmentation and Cup-Disc-Ratio (CDR) calculation on a reserved RIGA dataset. When tested on DRISHTI-GS and RIM-ONE dataset without re-training or fine-tuning, the model achieved comparable performance to that of the state-of-the-art in literature. We have also fine-tuned the model on two databases, which achieves an average disc dice value of 97.38% and cup dice value of 88.77% for DRISHTI-GS test set, disc dice of 96.10% and cup dice of 84.45% for RIM-ONE database, which is the state-of-the-art performance on both databases in terms of cup dice and disc dice value. The advantage of the proposed method is the combination of the pre-trained ResNet and U-Net, which avoids training the network from scratch, thereby enabling fast network training with less epochs, thus further avoids over-fitting and achieves robust performance.