A Robust Methodology to Detect Glaucoma Disease using Modified Deep Learning Principles with Image Processing Logics

Glaucoma is a leading cause of irreversible blindness, requiring early and accurate diagnosis to prevent vision loss. This study proposes a robust methodology for detecting glaucoma using a hybrid approach that integrates deep learning (Autoencoder) with a machine learning classifier (Multi-Layer Perceptron). The methodology utilizes image processing techniques, including noise reduction and data augmentation, to enhance the quality of retinal fundus images. Feature extraction was performed using Convolutional Neural Networks (CNNs) and Gabor filters, followed by dimensionality reduction through Principal Component Analysis (PCA) and wavelet decomposition. The model was tested on multiple publicly available datasets, including RIGA, HRF, ACRIMA, and Kaggle. The hybrid model achieved an accuracy of 97.1% on the RIGA dataset, 95.6% on the HRF dataset, and 96.3% on the ACRIMA dataset. The model outperformed traditional classifiers such as Support Vector Machine (SVM) and Random Forest, which achieved accuracies of 92.3% and 94.5% on the RIGA dataset, respectively. The study demonstrates that combining deep learning with machine learning classifiers, along with effective preprocessing, significantly improves glaucoma detection accuracy. Furthermore, explainable AI techniques were employed to enhance model interpretability, ensuring that clinical practitioners can trust and understand the model's decisions.