Abstract / Summary
Age-related macular degeneration (AMD) and cataract are among the leading causes of visual impairment worldwide. However, early diagnosis often depends on manual clinical assessment, which can be subjective and inconsistent. Automated analysis of retinal fundus images offers a scalable solution for reliable screening, but single deep learning models frequently suffer from overfitting and limited generalisation. This study proposes ARM-C, an ensemble deep learning framework designed to improve the robustness and reliability of retinal disease classification. The framework integrates five pre-trained convolutional neural networks, namely VGG16, VGG19, ResNet50, EfficientNetB0, and DenseNet121, whose predictions are combined using a majority voting strategy to reduce model variance and enhance classification stability. The proposed system was trained and evaluated on the publicly available ODIR dataset consisting of 4500 retinal fundus images across three categories: normal, cataract, and AMD. Experimental results demonstrate that ARM-C achieves an overall accuracy of 98%, with 98% sensitivity, 98% precision, and 99% specificity, outperforming several existing single-model approaches. The ensemble framework provides consistent decision behaviour across heterogeneous architectures, reducing diagnostic uncertainty and improving predictive robustness. To the best of our knowledge, this study is among the first to jointly classify AMD and cataract using a five-model ensemble optimised for automated screening. The results highlight the potential of ensemble deep learning to deliver reliable, cost-effective diagnostic support for ophthalmic screening in both specialised clinics and resource-limited healthcare environments. Nevertheless, the proposed framework has not yet been validated on external clinical datasets and does not incorporate explainable AI techniques, which will be explored in future work to enhance its generalizability and clinical interpretability.