A Hybrid Classification Approach for Glaucoma Diagnosis with MultiSURF Feature Selection

Abstract

Glaucoma is a severe optic neuropathy that can lead to irreversible vision loss and is often asymptomatic in its early stages. This situation increases the need for automated and reliable screening systems, especially those based on retinal fundus images. In this study, a hybrid classification framework suitable for clinical application and possessing high generalizability is presented, combining classical image processing methods with deep learning-based representations for glaucoma detection. In the proposed approach, deep features obtained from the fc7 layer of a pre-trained AlexNet architecture, along with HSV colour histograms, Haralick textural features, LBP and HOG descriptors, are combined using the ORIGA retinal fundus dataset. The class imbalance problem was addressed by applying the SMOTE method exclusively to the training data, while the high-dimensional feature space was optimised using the MultiSURF algo-rithm. The selected features obtained were evaluated using Support Vector Machines (SVM), Random Forest (RF), and XGBoost classifiers. Experimental results demonstrate that the proposed hybrid framework can accurately distinguish between individuals with glaucoma and those without, with high accuracy. Specifically, the Random Forest model demonstrated the highest performance with 96.92% accuracy, an F1 score of 0.939, and an AUC value of 0.994. The findings reveal that the combined use of deep and hand-crafted features, supported by MultiSURF-based feature selection and balanced learning strategies, offers a robust and stable solution for glaucoma detection. The proposed approach is considered a promising automated glaucoma screening method that can be integrated into clinical decision support systems for early diagnosis.