Early detection of retinal diseases like diabetic retinopathy, glaucoma, and cataract is crucial to avoid vision impairment. Deep learning has shown high diagnostic accuracy in fundus image classification, but most methods depend on large, labeled datasets, which restricts their use in data-limited settings. This study assesses a one-shot learning approach for classifying multiple retinal diseases with a Siamese neural network. The goal was to see if reliable performance can be sustained with just one labeled support image per class during inference. The dataset had four categories: cataract, diabetic retinopathy, glaucoma, and normal, split into stratified training and testing subsets. A shared-weight convolutional backbone was refined with contrastive loss to create a distinct embedding space. Classification involved comparing query and support embeddings based on distance. Performance was evaluated through accuracy, F1-score, confusion matrix, and one-vs rest ROC-AUC, utilizing Grad-CAM for interpretability. The method reached 92.77% accuracy and a macro-AUC of 0.979 in one-shot conditions. Results show that metric-learning enables strong multi-class discrimination and greatly cuts down on annotation needs.
Limitations are the dataset scope and the requirement for external clinical validation.