A Comparative Study on Few-Shot Learning for Retinal Disease Classification
Resumo
The scarcity of annotated data in medical domains poses a major challenge for training deep learning models. Motivated by this, we evaluated 14 Few-Shot Learning (FSL) methods for retinal disease classification on BRSET, a Brazilian ophthalmology dataset, encompassing both the CNN and Vision Transformer backbones. We further assessed explainability using Grad-CAM to ensure that model decisions are aligned with pathological features. Our results highlight the potential of meta-learning for retinal image analysis, demonstrating that few-shot methods can achieve robust performance with minimal annotated data while maintaining interpretability through alignment with clinically relevant features. To support reproducibility and future research, we present Simple FewShot, an open-source library. The library is available at github.com/victor-nasc/SimpleFewShot.
