LsEyeNet: deep feature fusion-based models for a large spectrum of eye disease recognition from ophthalmic images

Abstract

Early and accurate identification of eye diseases is crucial for avoiding vision loss and providing appropriate treatments. Several researchers focus on eye disease detection, but limited spectrum. In this study, we have proposed 9 disease classification fusionbased models for a large spectrum of eye disease recognition. Here, we have extracted attention-based depth level features and utilized the ResNet50 architecture for eye disease detection. Secondly, we have utilized ConvNeXtBase and EfficientNetB3 for the latent space features, then fused them to a fully connected neural network for large-scale spectrum eye disease classification. To evaluate the performance of our proposed models, we have utilized the Eye Disease Image Dataset (Mendeley Data) and obtained superior accuracy. We have compared model performances with existing models and self-supervised approaches. The hybrid LsEyeNet model performed the solo and baseline models, achieving 87.37 percent accuracy, high precision (0.89), strong recall (0.87), and the highest F1-score (0.89). These results demonstrate the efficacy of mixing contemporary convolutional architectures for accurate eye disease categorization. Furthermore, our findings reveal that, while self-supervised learning using SimCLR falls short of fully supervised techniques in a fully labeled context, it remains a promising strategy when annotated data is insufficient. This comprehensive performance benchmark highlights hybrid models’ potential for developing trustworthy AI-based diagnostic tools in ophthalmology.