Dual-Level Adaptive Sampling for Enhanced Few-Shot Medical Image Classification

Medical image classification is often hindered by limited labeled data, making few-shot learning (FSL) a crucial approach. However, conventional FSL methods struggle with the high variance present in medical images. Consequently, we propose Dual-Level Adaptive Sampling (DLAS) to enhance FSL by strategically selecting the most informative class combinations and instances within each class. During training, our class sampling policy prioritizes challenging, underconfident classes to foster continual learning. Subsequently, instances within each class are weighted according to their true class probability, ensuring the selection of hard-to-classify samples for improved generalization. This dual-level strategy is seamlessly integrated into prototypical networks within a metric-based meta-learning framework and can be adapted to various few-shot learning models. We validate our approach through extensive quantitative and qualitative analyses on three benchmark medical imaging datasets-Derm7pt, PathMNIST, and ChestMNIST. Results show that our sampling framework significantly improves model generalization and classification performance compared to existing State-of-the-art methods.