Comparative Analysis and Validation of Performance Metrics for Ocular Disease Detection using Deep Learning Models

It is often challenging for health experts and low-experienced ophthalmologists to accurately diagnose ocular diseases in their early stages. Manual diagnosis is inefficient, error-prone, and complicated due to the presence of multiple eye conditions like as diabetic retinopathy, glaucoma, cataract, and other retinal abnormalities. To overcome these challenges, an automated and intelligent Computer-Aided Diagnostic (CAD) system is essential for reliable detection of ocular diseases. In recent years, Deep Learning (DL) strategies have greatly advanced in enhancing the classification performance of medical imaging systems. In this paper, we present a comparative study focused on evaluating an optimized detection method based on deep feature extraction and selection strategies for classifying eye-related disorders. The proposed system, HFBOCNN (Hybrid VGG-16 Feature-Based Optimized CNN), employs a pre-trained VGG-16 network for deep feature extraction, followed by an intelligent feature selection procedure by utilizing the Crow Search Algorithm (CSA). These refined features are then utilized to identify and classify ocular illness conditions. To validate the effectiveness of our approach, HFBOCNN is compared with several existing DL models, including VGG-16, ResNet-50, and the AI-GS Network. The experiments were conducted using benchmark ocular datasets, and standard data augmentation techniques were applied to manage class imbalance. Among all tested models, HFBOCNN achieved the best classification accuracy of 96.47% and the lowest error rate of 3.53%. These results highlight the significance of combining deep features with optimized feature selection techniques to enhance disease detection performance.