Meta-Learning Based Few-Shot Classification of Retinal Diseases

To address sample scarcity, a common challenge in medical imaging, we investigate the Reptile meta-learning algorithm for few-shot disease classification in fundus images. In a setup with eight training classes and four testing classes, we investigate different architectures, training strategies, varying N-way K-shot configurations, and different data augmentation techniques. Our results show that Reptile outperforms standard transfer learning approaches in several settings and that, when combined with data augmentation, especially during evaluation, correct predictions tend to receive higher confidence scores. The quantitative results are further supported by an analysis of prediction confidence levels and activation-map visualizations.