Application of AlexNet, EfficientNetV2B0, and VGG19 with Explainable AI for Cataract and Glaucoma Image Classification

The rapid integration of Artificial Intelligence (AI) has significantly improved healthcare outcomes, especially in ophthalmology. However, Deep learning algorithms are often called to as 'black-box' models, making interpreting decision processes difficult. The Explainable AI approach helps understand AI model predictions by explaining the decision-making process based on features. The research objective is to find the best model for classifying eye disease images through experiments using AlexNet, EfficientNetV2, and VGG19 architectures. The selected best model will be analyzed using Grad-CAM, enhancing transparency and interpretability in healthcare AI models. This study employs multi-class classification using AlexNet, EfficientNetV2B0, and VGG19 architectures to classify cataract, glaucoma, and normal fundus. Different holdout-validation techniques (70:30, 80:20, 90:10) and optimization strategies (Adam, AdamW, RMSProp, SGDM) are studied for effective disease classification. The experimental results found that the best-performing model, EfficientNetV2B0 with holdout 90:10 and RMSProp optimizer, outperformed other models in accurately classifying fundus images. The model demonstrated strong performance with high recall, precision, f1-score of 0.8969, 0.8970, 0.8969, and an accuracy rate of 89.77%. Grad-CAM analysis also identified unique features of each eye disease class, for example, cataracts. AI model is capable of detecting cataract characteristics through the level of opacity in the eye lens, underscoring its robust classification capabilities in eye disease diagnosis.