Unsupervised Representation Learning Meets Pseudo-Label Supervised Self-Distillation: A New Approach to Rare Disease Classification

Rare diseases are characterized by low prevalence and often chronically debilitating or life-threatening. Imaging-based classification of rare is challenging due to the severe shortage in training examples. Few-shot learning (FSL) methods tackle this challenge extracting generalizable prior knowledge from a large base dataset common normal controls, transferring diseases. Yet, most existing require be labeled do not make full use precious examples To end, we propose work novel hybrid approach disease classification, featuring two key novelties targeted at above drawbacks. First, adopt unsupervised representation (URL) based on self-supervising contrastive loss, whereby eliminate overhead labeling dataset. Second, integrate URL with pseudo-label supervised for effective self-distillation about diseases, composing taking advantages both (pseudo-) Experimental results skin lesions show that our substantially outperforms FSL (including those using fully dataset) via integration driven self-distillation, thus establishing new state art.