Enhancing Unsupervised Few-Shot Medical Image Classification with Weight-Enhanced Contrastive Learning

In the realm of contemporary deep learning, the pivotal challenges confronting few-shot and unsupervised learning reside in harnessing scarce labeled samples alongside abundant unlabeled ones, particularly in the context of medical image classification. This paper introduces an innovative approach that seamlessly integrates dynamic clustering with weights augmentation, aimed at bolstering the performance of few-shot medical image classification. Dubbed Weight-Enhanced Contrastive Learning (WECL), our method ingeniously fuses contrastive representation learning with a dynamic memory module during unsupervised pre-training. This fusion facilitates efficient clustering and classification of diverse augmented renditions of the same image. Additionally, the weights augmentation tactic meticulously tunes the weights of both ResNet and teacher-student model branches, thereby mitigating sample bias and enhancing the pre-trained model's proficiency. Extensive experiments across multiple few-shot medical image classification datasets underscore the superiority of our WECL approach, outperforming current state-of-the-art baselines, and effectively addressing issues pertaining to data distribution disparities and sample scarcity.