Artificial intelligence-driven glaucoma screening in ophthalmology: The GlaucoNet deep learning framework

Abstract

Background: Glaucoma is a significant cause of permanent blindness in the world, and it is usually unnoticed at its initial stages. Deep learning allows scaling flexible quality screening of retinal fundus images to detect and implement effective intervention at the appropriate time.
Aim: This study developed and evaluated GlaucoNet, an ensemble of MobileNetV2, InceptionV3, and ResNet50, for accurate glaucoma detection from retinal fundus images.
Setting: Publicly available data sets used include Online Retinal fundus Image database for Glaucoma Analysis (325 normal and 325 glaucoma images), and Automatic Classification of Retinal Images for Medical Assessment (309 normal and 396 glaucoma images).
Methods: To address the imbalance in classes, the process of data augmentation such as flipping, rotation, scaling and brightness was used. The data were separated into 5-fold cross-validation with patient-wise data separation training (70%), validation (10%) and testing (20%) to avoid data leakage. The early stopping that was done on validation loss, dropout (0.3) and L2 regularisation minimised overfitting in the branches of the convolutional neural network (CNN).
Results: The accuracy of GlaucoNet was 97.0, sensitivity 97.32, and specificity 96.64, with the following 95% confidence intervals of 96.19–97.8, 96.39–98.68 and 95.59–96.63 respectively. DeLong testing showed that the area under the curve of GlaucoNet (0.9698) was significantly greater compared to MobileNetV2, InceptionV3 and ResNet50 (p < 0.05) and proved the statistical dominance of the ensemble method.
Conclusion: GlaucoNet achieved high accuracy in glaucoma detection, supporting its clinical screening potential.
Contribution: By leveraging the strengths of multiple CNN architectures, GlaucoNet offers a promising tool for enhancing early diagnosis and facilitating timely clinical intervention in ophthalmology.