Leveraging lightweight vision architecture for effective multiclass eye disease detection in fundus photographs

Most existing eye disease detection models focus on binary classification and exclude several vision-threatening conditions. This study proposes a lightweight deep learning model to classify nine retinal conditions, including clinically important diseases absent from benchmark datasets like ODIR-5K, using color fundus images. While high accuracy models are widely used in specialized hospitals, their reliance on substantial computational resources limits deployment in mobile screening programs, rural clinics, or resource-limited settings. Therefore, this study explores models that balance predictive performance and computational efficiency to support early screening in such environments. Rather than replacing clinical judgment, the model is designed to assist healthcare professionals by flagging potential abnormalities and streamlining the diagnostic workflow. Four lightweight architectures, which are MobileNetV4, LeViT, EfficientViT, and GENet were compared on a dataset consisting of 21,000 expert-annotated images. EfficientViT achieved the best performance with PR-AUC 0.8874, ROC-AUC 0.9780, and Kappa 0.7899 despite having only 3M parameters. While the model performs well overall, detecting early-stage glaucoma is still challenging due to its subtle features. Additionally, coexisting diseases like glaucoma and retinal conditions can produce overlapping features, leading to misclassifications that reflect clinical complexity rather than model errors. Future works should explore multi-label classification and additional classes to better capture overlapping conditions.