Skin lesion classification with mini-batch sampling and deep metric learning

Skin lesion image classification based on deep learning has recently garnered significant attention. However, directly applying methods that perform well in general computer vision tasks to skin lesion image classification is not ideal, as skin lesion image datasets possess intrinsic characteristics, such as class imbalance, intra-class variability, and inter-class similarity. To tackle these challenges simultaneously, we propose a novel unified learning framework, named mBSML, which integrates mini-batch sampling and deep metric learning. In this framework, mini-batch sampling re-samples data in real-time during each iteration of learning, while a new loss function combines mini-batch distance metric-based loss with cross-entropy loss. Through the alternating training procedure on both imbalanced training data and balanced re-sampling data, mBSML effectively learns from global distribution information and local similarity information, not only from the original dataset but also from the minority classes. Extensive experiments conducted on two publicly available datasets demonstrate the effectiveness of mBSML for skin lesion image classification.