Hybrid Deep Learning Architecture for Glaucoma Detection: Integrating a Multi-Network CNN Ensemble with ANFIS

Abstract

Glaucoma is a leading cause of irreversible blindness worldwide, making early and accurate diag¬nosis critical for preventing severe vision loss [1]. While deep learning has advanced computer-aided diagnostics, traditional single-stream Convolutional Neural Networks (CNNs) [2] often strug¬gle with overfitting on small datasets and managing diagnostic uncertainties in overlapping disease patterns. To address these limitations, this thesis proposes a novel hybrid architecture that inte¬grates a multi-network CNN ensemble with an Adaptive Neuro-Fuzzy Inference System (ANFIS), as implemented and evaluated in the accompanying Python programmatic framework. Specifi¬cally, the Python pipeline pre-processes retinal fundus images via Region of Interest (ROI) ex¬traction and enhancement, feeding them into a parallel feature extraction module utilizing pre-trained ResNet, DenseNet, and MobileNet backbones. These fused, high-level structural features are subsequently passed into the ANFIS module, which applies adaptive fuzzy reasoning to ef¬fectively manage diagnostic uncertainty and subtle structural variations. Evaluated on aggregated benchmark datasets—including RIM-ONE, DRISHTI-GS1, and ACRIMA—the custom MultiNet-ANFIS Python program demonstrates superior diagnostic performance. When directly compared to standalone baseline models (such as standard ResNet18, DenseNet121, and MobileNet) within the script’s testing loop, the proposed CNN-ANFIS framework achieves significantly higher Ac¬curacy, Precision, and Area Under the Curve (AUC). By synergizing the robust feature extraction capabilities of ensemble deep learning with the interpretable decision-making of fuzzy logic, this programmatically validated model successfully mitigates classification errors and offers a highly efficient, scalable solution for automated glaucoma screening.