A proxy-enhanced and margin-adaptive contrastive learning method for skin lesion diagnosis

Deep neural networks (DNNs) have shown remarkable success in skin disease diagnosis, but their deployment is limited by data imbalance. This imbalance arises not only from the unequal class sample sizes in the training set, but also from the varying diagnostic difficulty among skin lesions. To address these challenges, we propose an innovative Proxy-Enhanced and Margin-Adaptive Contrastive Learning (PMCL) framework that incorporates adaptive margin and proxy mechanisms. Firstly, we develop an Integrated Difficulty Coefficient to dynamically assess class difficulty and adaptively emphasize hard classes during training. Secondly, we design a Class-Balanced Adaptive Margin Loss (CBAML) and a Class-Balanced Progressive Attention Strategy, introducing the integrated number ratio coefficient and integrated difficulty ratio coefficient to control class margins, dynamically adjusting the two contribution weights of the coefficients in CBAML to progressively shift attention from sample-size imbalance to class difficulty imbalance, fully considering their impact on classification performance and thereby enhancing the model’s representation ability while mitigating class imbalance. Finally, we propose a Class-Balanced Contrastive Learning Loss (CBCLL) and a Class-Balanced Proxy Generation Module, which allocates more learnable proxies to hard classes and minority classes, and fewer proxies to majority classes and easy classes, ensuring equal contribution of all classes to model training. We conducted extensive experiments on two datasets and achieved the best classification results, including a mean sensitivity of 85.01% and an accuracy of 91.41% on ISIC2018, as well as a mean sensitivity of 82.92% and an accuracy of 88.29% on ISIC2019. Compared with recent mainstream contrastive learning methods (e.g., ECL), PMCL achieves a 3.82% higher mean sensitivity and a 1.99% higher accuracy on the ISIC2018 dataset, as well as a 2.34% higher mean sensitivity and a 1.32% higher accuracy on the ISIC2019 dataset. Code is available in https://github.com/FXDF1004/PMCL . • We propose a Proxy-Enhanced and Margin-Adaptive Contrastive Learning (PMCL) framework to jointly address sample-size imbalance and difficulty imbalance in skin disease diagnosis. An Integrated Difficulty Coefficient is introduced to dynamically assess class-wise difficulty and adaptively emphasize hard classes during training. • We design a Class-Balanced Adaptive Margin Loss(CBAML) with a Progressive Attention Strategy to address sample-size and difficulty imbalance. In addition, a Class-Balanced Contrastive Learning Loss(CBCLL) with a Proxy Generation Module introduces difficulty-aware learnable proxies to enrich hard-class representations. • Extensive experiments on two large public datasets validate that PMCL outperforms existing methods and achieves state-of-the-art results.