An Efficient Glaucoma Prediction Using an Optimized CNN and XGBoost with Attention-Based Deep Learning Framework

Abstract

Over the Decade, The machine learning incorporated on image classification has gained increasing attention for extracting discriminative patterns from retinal fundus images and supporting automated glaucoma severity analysis respectively. The several studies have explored convolutional neural networks, feature extraction, support vector machines, Random Forest, and ensemble learning strategies for glaucoma classification. The study proposes a novel weighted ensemble machine learning framework integrating LightGBM for glaucoma severity classification. The proposed framework begins with preprocessing of retinal fundus images using contrast enhancement, normalization, and noise reduction. Multiple feature representations, including structural, texture, morphological, and CNN-based deep features, are generated to capture complementary characteristics of the input images. The features are provided to multiple machine learning classifiers to obtain individual class-probability predictions respectively. A performance-based weighted ensemble mechanism is needed to assign differentiated weights to the classifier outputs according to their predictive contribution. Providing an enhanced computational framework for automated glaucoma severity classification respectively.