Class-aware augmentation contrastive learning for long-tailed medical image classification

Although supervised contrastive learning has achieved success in long-tailed medical image classification, existing methods suffer from insufficient learning of tail classes and bias towards head classes, resulting in biased diagnostic models. This paper proposes a Class-aware Augmentation Contrastive Learning (CACL) method for long-tailed medical image classification, which enriches the information of each class by feature augmentation. Firstly, the distribution of the augmented anchors of a sample anchor is modeled as a Gaussian distribution centered at that sample anchor. To simplify the process of sampling augmented anchors via Gaussian distribution, we derive an upper-bound loss function of the contrastive loss, called balanced implicit augmentation contrastive loss (BIACL). Secondly, we propose the Balanced Hybrid Contrastive Loss (BHCL) to mine the additional information between sample anchors and learnable category prototype, and balance the contribution of all classes and ensure the fairness of learning. Finally, to enhance the classification performance on both head and tail classes thereby efficiently aiding medical diagnosis, we propose an Adaptive Re-balanced Cross-Entropy loss (ARCE). Extensive experiments demonstrate that the proposed CACL outperforms the existing state-of-the-art methods on three benchmark datasets.