Unleashing the potential of open-set noisy samples against label noise for medical image classification

Addressing the coexistence of closed-set and open-set label noise in medical image classification remains a largely unexplored challenge. Unlike natural image classification, where noisy samples can often be clearly separated from clean ones, medical image classification is complicated by high inter-class similarity, which makes the identification of open-set noisy samples particularly difficult. Moreover, existing methods typically fail to fully exploit open-set noisy samples for label noise mitigation, either discarding them or assigning uniform soft labels, thus limiting their utility. To address these challenges, we propose the ENCOFA: the Extended Noise-robust Contrastive and Open-set Feature Augmentation framework for medical image classification. This framework introduces the Extended Noise-robust Supervised Contrastive Loss, which enhances feature discrimination across both in-distribution and out-of-distribution classes. By treating open-set noisy samples as an extended class and weighting contrastive pairs based on label reliability, this loss effectively improves the robustness to label noise. In addition, we develop the Open-set Feature Augmentation module, which enriches open-set samples at the feature level and dynamically assigns class labels, thereby leveraging model capacity while mitigating overfitting to noisy data. We evaluated the proposed framework on two synthetic noisy datasets and one real-world noisy dataset. The results demonstrate the superiority of ENCOFA over six state-of-the-art methods and highlight the effectiveness of explicitly leveraging open-set noisy samples in combating label noise. The code will be publicly available at https://github.com/Merrical/ENCOFA.