.png){kind=icon}
.png){kind=icon}
Return
Multi-level temporal stabilization for semi-supervised skin lesion classification
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neucom.2026.134596.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Semi-supervised learning (SSL) offers a practical solution for dermoscopic image classification, where expert annotations are limited and costly. However, existing SSL methods often suffer from confirmation bias due to noisy pseudo-labels and from augmentation-induced noise that distorts clinically relevant lesion features. To address these limitations, we propose Multi-Level Temporal Stabilization (MLTS), a unified SSL framework that stabilizes learning in both prediction and representation spaces.MLTS adopts a teacher–student paradigm in which the teacher is updated via exponential moving average of the student parameters and generates temporally ensembled predictions across iterations. These aggregated predictions produce stable pseudo-labels that mitigate confirmation bias. At the feature level, we introduce per-image prototypes computed through exponential averaging of historical augmented embeddings, providing low-variance, instance-specific targets that suppress augmentation artifacts while preserving discriminative lesion structure. A two-branch student architecture further decouples invariance learning from classification. Extensive evaluation across four benchmark datasets (ISIC 2018, ISIC 2019, PH2, and PAD-UFES-20) demonstrates that our method consistently outperforms fully supervised and state-of-the-art SSL baselines. These improvements span both dermoscopic and clinical smartphone modalities, proving particularly robust under severe label scarcity and class imbalance.The proposed approach achieves substantial gains in balanced accuracy and macro-F1, confirming the effectiveness of MLTS for reliable skin lesion classification. Code is publicly available at: https://github.com/skrishnan19/MLTS .
Journal
IF:
6.5
Papers:
2.5W
Citations:
6.5W
View more organizations