Reliable Contrastive Pseudo-Labels Network for Unsupervised Skin Lesion Segmentation

Unsupervised skin lesion segmentation offers significant advantages by eliminating the dependency on large annotated datasets, a common challenge in medical image analysis. However, traditional methods often struggle with accuracy and robustness. Recent advances utilizing pseudo-labels instead of ground truth have shown potential, but improving pseudo-label quality and mitigating the impact of label noise remain critical challenges. In this work, we propose a novel approach, the Reliable Contrastive Pseudo-Labels Network for Unsupervised Skin Lesion Segmentation. Our method leverages multiple contrastive learning to extract image features, followed by clustering to generate pseudo-labels. To address label noise, we introduce a multi-source uncertainty mining strategy that analyzes the commonalities and discrepancies across multiple pseudo-labels, generating uncertainty maps. This effectively reduces the influence of noisy labels while enhancing interpretability. Experiments on the ISIC-2017 dataset demonstrate that our method achieves significantly improving of segmentation accuracy.