Attention-Enhanced Prototypical Networks for Few-Shot Microaneurysm Detection in Diabetic Retinopathy Images

Early detection of microaneurysms in diabetic retinopathy (DR) is essential for prohibiting vision loss, yet automated detection remains difficult due to the limited labeled data and the subtle nature of these lesions. This paper presents a few-shot learning approach that synergistically combines dual attention mechanisms with prototypical networks for microaneurysm detection in fundus images. Our architecture implements a modified ResNet-50 backbone enhanced with spatial and channel attention modules, optimized specifically for small lesion detection. The spatial attention module generates precise localization maps, while the channel attention mechanism adaptively weights feature importance. This is integrated with a prototypical network framework that enables efficient learning from limited labeled examples through metric-based classification. Evaluated on the IDRiD dataset, our method achieves state-of-the-art performance with an AUC-ROC of 0.947, sensitivity of 0.892, and specificity of 0.941, significantly outperforming traditional deep learning approaches. The model demonstrates robust performance with as few as 5 shots per class, while maintaining high detection accuracy (0.915 F1-score) using only 5% of the training data needed by conventional methods. Ablation studies demonstrate the effectiveness of our dual attention mechanism, with the full model showing a 2.4% improvement in AUC-ROC over the baseline prototypical network.