<p>Deep neural networks (DNNs) have achieved remarkable success in medical image classification tasks, but the performance of DNN-based methods in medical scenarios remains limited. Sample weighting has shown substantial promise in addressing this issue by assigning weights for each sample to assess its importance. However, existing methods rely on individual samples for weighting and neglect the relationships among samples, which limits their potential in medical image classification. To this end, we propose a new Class Correlation-based Dynamic Sample Weighting (CC-DSW) framework for medical image classification. CC-DSW model sample relationships both in feature and label space by leveraging the intra-class and inter-class correlations, capturing intra-class consistency and inter-class separability. It then maps the class correlations to the sample weights through a learnable sample weighting network, allowing for automatic weight assignment during training. The sample weighting network and the task network are optimized alternately using meta-learning for mutual adaptation. We evaluate the effectiveness of our method on three medical image classification benchmarks: PatchCamelyon for lymph node histopathology classification, ISIC 2020 for skin lesion classification, and MTC for medullary thyroid carcinoma classification. CC-DSW outperforms existing state-of-the-art sample weighting methods across all three datasets and significantly exceeds methods without sample weighting. Compared with the ACC, F1 and AUC of the baseline, our proposed CC-DSW improves by 5.00%, 4.61% and 3.23% in PCam, 2.30%, 1.81% and 2.62% in ISIC 2020, and 5.63%, 6.88% and 3.55% in MTC. Experimental results demonstrate that CC-DSW leverages class correlation for dynamic weighting, which makes the model focus on samples at the decision boundary and improves the performance in medical image classification tasks.</p>