Advancing Eye Disease Assessment through Deep Learning: A Comparative Study with Pre-Trained Models

The significant global challenges in eye care are treatment, preventive quality, rehabilitation services for eye patients, and the shortage of qualified eye care professionals. Early detection and diagnosis of eye diseases could allow vision impairment to be avoided. One barrier to ophthalmologists when adopting computer-aided diagnosis tools is the prevalence of sight-threatening uncommon diseases that are often overlooked. Earlier studies have classified eye diseases into two or a small number of classes, focusing on glaucoma, and diabetes-related and age-related vision issues. This study employed three well-established and publicly available datasets to address these limitations and enable automatic classification of a wide range of eye disorders. A Deep Neural Network for Retinal Fundus Disease Classification (DNNRFDC) model was developed, evaluated based on various performance metrics, and compared with four established pre-trained models (EfficientNetB7, EfficientNetB0, UNet, and ResNet152) utilizing transfer learning techniques. The results showed that the proposed DNNRFDC model outperformed these pre-trained models in terms of overall accuracy across all three datasets, achieving an impressive accuracy of 94.10%. Furthermore, the DNNRFDC model has fewer parameters and lower computational requirements, making it more efficient for real-time applications. This innovative model represents a promising avenue for further advancements in the field of ophthalmological diagnosis and care. Despite these promising results, it is essential to acknowledge the limitations of this study, namely the evaluation conducted by using publicly available datasets that may not fully represent the diversity and complexity of real-world clinical scenarios. Future research could incorporate more diverse datasets and explore the integration of additional diagnostic modalities to further enhance the model's robustness and clinical applicability.