An ensemble framework based on Deep CNNs architecture for glaucoma classification using fundus photography.

Area of Science:

• Ophthalmology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computer Vision for Disease Detection

Background:

• Glaucoma is a leading cause of irreversible blindness.
• Early detection and classification are crucial for effective treatment.
• Deep Convolutional Neural Networks (Deep CNNs) have shown potential in medical image analysis.

Purpose of the Study:

• To propose and evaluate a two-stage glaucoma classification system using Deep CNNs.
• To compare the performance of various pre-trained Deep CNN architectures.
• To develop novel ensembling techniques to improve classification accuracy.

Main Methods:

• Utilized four ImageNet pre-trained Deep CNNs (AlexNet, InceptionV3, InceptionResNetV2, NasNet-Large) in the first stage.
• Evaluated classifier performance on public (ACRIMA, ORIGA-Light, RIM-ONE) and local (AFIO, HMC) datasets.
• Developed a two-stage ensemble classifier using accuracy-based weighted voting and accuracy/score-based weighted averaging.

Main Results:

• NasNet-Large demonstrated superior performance as a single classifier (Sensitivity: 99.1%, Specificity: 99.4%, Accuracy: 99.3%, AUC: 97.8%).
• The proposed ensemble classifier, particularly accuracy/score-based weighted averaging, further improved classification accuracy to 99.5%.
• Consistent high performance was observed across diverse public and local datasets.

Conclusions:

• NasNet-Large is a highly effective single-model solution for automatic glaucoma classification.
• Ensemble methods, especially accuracy/score-based averaging, significantly enhance generalized glaucoma classification performance.
• The proposed two-stage approach offers a robust and accurate method for glaucoma detection in clinical settings.