Few-shot learning for rare skin disease classification via adaptive distribution calibration

The classification of rare skin diseases faces significant data scarcity challenges due to the difficulty in acquiring clinical samples and the high cost of annotation, which severely hinders the training of deep neural network-based models. Few-shot learning has emerged as a cutting-edge solution, with its core capability being the identification of novel disease classes using limited annotated samples to mitigate data insufficiency. However, most existing methods fail to fully leverage the statistical information from base classes to calibrate the distribution of few-shot classes, thereby optimizing classifier inputs. Two critical research challenges remain: (1) accurately estimating the true distribution of few-shot classes with minimal samples, and (2) selecting appropriate base class information for effective distribution calibration. To address these challenges, we propose SADC (skin disease classification via adaptive distribution calibration), a new few-shot learning framework incorporating multi-scale feature extraction and adaptive sample calibration. First, our multi-scale feature extraction strategy employs feature descriptor matrices and composite metrics to optimize multi-dimensional, multi-directional feature representations, enabling precise similarity computation between base-class and few-shot samples. Second, the adaptive sample calibration strategy constructs weight matrices based on sample similarity to automatically select optimal base-class samples with adaptive weights for distribution calibration, ensuring alignment between calibrated distributions and true unbiased distributions. Experimental results demonstrated that SADC achieves state-of-the-art performance across three public dermatology datasets (ISIC2018, Derm7pt, and SD198), showing significant improvements over existing methods. The framework's innovation lies in its dual-strategy approach to distribution-aware few-shot learning, advancing the frontier of data-efficient medical image analysis.