Hybrid Few-Shot Learning With Attention for Multi-Disease and Out-of-Distribution Detection in Retinal Imaging

Early and accurate detection of retinal diseases is crucial for preventing vision loss, especially in resource-limited settings where labeled data is scarce. In this paper, a hybrid few-shot learning framework for multidisease detection and out-of-distribution (OOD) identification in retinal imaging is presented. The framework integrates sparse cross-attention and multi-attention mechanisms to enhance feature extraction by focusing on both global and local image regions, allowing the model to capture subtle disease-related patterns. Sparse attention reduces the computational cost by 26.8% (from 4.1 to 3.0 GFLOPs) while maintaining competitive classification performance. Prototype-based metric learning is used to facilitate a few-shot classification, allowing the model to effectively recognize diseases with limited labeled samples. The results demonstrate that the model is capable of distinguishing multiple retinal diseases, handling class imbalances, and identifying out-of-distribution samples, positioning it as a practical solution for real-world deployment in ophthalmology, particularly in resource-aware clinical settings where computational efficiency relative to other attention-based models and the ability to flag atypical cases for expert review are critical.