Confidence-Weighted Semi-Supervised Learning for Skin Lesion Segmentation Using Hybrid CNN-Transformer Networks

Automated skin lesion segmentation through dermoscopic analysis is essential for early skin cancer detection; however, it remains challenging because of the limited annotated training data. We present MIRA-U, a semi-supervised framework that combines uncertainty-aware teacher-student pseudo-labeling with a hybrid CNN-Transformer architecture. Our approach employs a teacher network pretrained via masked image modeling to generate confidence-weighted soft pseudo-labels, which guide a U-shaped CNN-Transformer student network featuring cross-attention skip connections. This design enhances the pseudo-label quality and boundary delineation, surpassing the reconstruction-based and CNN-only baselines, particularly in low-annotation regimes. Extensive evaluation on the ISIC-2016 andPH2datasets demonstrates competitive performance, with a Dice Similarity Coefficient (DSC) of 0.9153 and Intersection over Union (IoU) of 0.8552 using only 50% labeled data, comparable to fully supervised methods using 100% of the labels. Code is publicly available at GitHub.