Dynamic learning of sample ambiguity-driven sample weighting for medical image classification

Deep neural networks have delivered impressive results in medical image classification tasks. However, their performance is still challenging in medical scenarios with limited data, where training set biases such as label noise or class imbalance impede model learning. Dynamic learning based sample weighting achieves adaptive adjustment of sample importance through learnable weight functions and shows great potential in improving model robustness. Nevertheless, existing methods directly employ model states such as loss value or training accuracy to evaluate sample importance, ignoring the role of ambiguous samples in model optimization. This limitation hinders the performance of dynamic learning based sample weighting in medical image classification. In this paper, we propose a new sample weighting approach based on sample ambiguity and dynamic learning for improving medical image classification, named DLSA-SW. We introduce a dual-space sample ambiguity method by evaluating the category proximity in the feature space and the prediction confidence in the label space. Subsequently, to dynamically calculate sample weights according to sample ambiguity, a learnable sample weighting network is developed to adaptively adjust the weights during training to guide the task model. DLSA-SW performs alternate optimization to enable mutual adaptation of the sample weighting network and the task network. We evaluate the effectiveness of our approach on three medical image classification benchmarks: PatchCamelyon for lymph node histopathology classification, ISIC 2020 for skin lesion classification, and MTC for medullary thyroid cancer classification. DLSA-SW outperforms existing state-of-the-art sample weighting methods on all three datasets and yields substantial improvements over methods without sample weighting. These results demonstrate the robustness and practical applicability of our approach in clinical diagnostic tasks.