Glaucoma, a principal cause of irreversible blindness, involves progressive optic nerve damage often linked to increased intraocular pressure. Traditionally, diagnosing glaucoma involves ophthalmologists examining a dilated pupil, a method that is both time-consuming and labor-intensive. Automation through deep learning offers a scalable solution, streamlining the diagnostic process. Convolutional Neural Networks (CNNs) excel in this context as they extract hierarchical features from images, enabling the distinction between glaucomatous and non-glaucomatous patterns crucial for accurate diagnoses. This study introduces a deep learning-based system for automated glaucoma detection, emphasizing the optic disc and cup segmentation through a dual-network architecture. The model employs two hybrid CNNs for simultaneous segmentation tasks, where each CNN outputs a mask for either the optic cup or the optic disc. These masks are merged into one image to produce the final segmentation result. We evaluate our model using three benchmark datasets: RIM-ONE, ACRIMA, and Drishti-GS, achieving glaucoma prediction accuracy of 96.42%, 93.2%, and 94.4%, respectively. Using the Dice coefficient, our model achieves cup/disc segmentation scores of 90/97% on RIM-ONE, 90/95% on ACRIMA, and 94/97% on Drishti-GS. These results are compared with recent methods, demonstrating the robustness and effectiveness of our approach across varying architectures. The pipeline is transparently interpretable: predicted optic cup/disc contours are visualized and the per-case cup-to-disc ratio (CDR) that drives the glaucoma decision is reported. This study is model-centric, emphasizing a hybrid architecture (ResNet-50 screening + Z-Net segmentation) evaluated on public datasets. The model and datasets are available here.
Algaraghuli et al. (Fri,) studied this question.