Adaptive Dynamic Contrastive Learning with Hard Negative Mining for Imbalanced Medical Image Classification

Abstract

Class imbalance poses a fundamental challenge in medical image classification, where rare pathologies are severely underrepresented in training datasets. This paper presents ADCL-DHNM (Adaptive Dynamic Contrastive Learning with Dynamic Hard Negative Mining), a novel framework that synergistically combines supervised contrastive learning with an adaptive hard negative mining strategy. Our approach introduces a class-aware contrastive loss with dynamic threshold adaptation based on training progress and class distribution, ensuring proportional representation of minority classes during feature learning.
We provide theoretical analysis demonstrating that our adaptive threshold mechanism approximates optimal sampling distributions and that the proposed loss function increases gradient contributions from minority classes. Extensive experiments across four benchmark medical imaging datasets (ISIC 2019, ChestX-ray14, HAM10000, and MIMIC-CXR) demonstrate state-of-the-art performance with AUC improvements of 8.2-9.1% over cross-entropy baselines. Rigorous statistical validation (paired t-tests, effect size analysis, bootstrap confidence intervals) confirms the significance of our results (p < 0.001). Comprehensive stress testing under extreme imbalance (up to 1000:1), label noise (up to 30%), and limited data scenarios validates the robustness of our approach.