Enhanced hard negative synthesis for contrastive learning in skin lesion classification

The analysis of skin lesion dermoscopic images has traditionally relied on supervised learning paradigms, necessitating extensive labeled datasets which can be costly and time-consuming for dermatologists. Conversely, the abundance of unlabeled data offers a promising avenue for self-supervised learning methods in this field. In this work, we introduce a novel strategy for hard negative synthesis that bolsters the efficacy of contrastive learning in skin lesion classification. By strategically down-weighting the contribution of the hardest negative samples during feature-level mixing, our method ensures the neural network prioritizes learning from the most informative and reliable negatives, thereby enhancing the model's feature learning ability. After fine-tuning with a limited set of labeled data, our method demonstrates notable superiority on the ISIC-2016 classification dataset, achieving a 4.29% increase in ROC AUC and a 5.39% increase in F1 score over the standard MoCo-v2 framework.