GAMFuse: Graph-based Adaptive Multiscale Feature Fusion for few-shot skin lesion classification

Automatically classifying skin diseases is crucial for enhancing diagnostic accuracy, especially when labeled data is scarce. Few-shot learning techniques, notably Graph Neural Networks (GNNs), offer substantial promise by leveraging relational reasoning and knowledge aggregation. However, the significant intra-class variation and inter-class similarity in skin lesions complicate the extraction of robust and discriminative features, often hindering classification performance. To address these issues, we propose the Graph-based Adaptive Multiscale Feature Fusion (GAMFuse) framework. GAMFuse effectively integrates multiscale global context and local details of lesions across three parallel branches: a global branch utilizes a Swin Transformer to capture long-range interactions and high-level semantics, a local branch with a shallow residual convolutional network extracts fine-grained texture and structural details, and a Multisource Attention Fusion (MAF) branch adaptively fuses these features through a self-attention mechanism. This approach ensures strong feature representations while preventing interference between branches. The fused features are further processed through Graph Neural Networks (GNNs) , which model inter-lesion relationships, enhancing classification accuracy, especially in few-shot learning scenarios. In a 2-way 5-shot setting, our approach achieves accuracy rates of 97.41%, 95.61%, and 92.50% on the SD-198, Derm7p, and ISIC2018 datasets, respectively, surpassing state-of-the-art approaches.