Teacher–Student Pseudo-Labeling for Semi-Supervised Skin Lesion Segmentation: A Reproducible Baseline

Obtaining expert-annotated segmentation masks for medical images is expensive and time-consuming, limiting the scalability of supervised deep learning approaches in clinical settings. We present a systematic study of teacher–student pseudo-label semi-supervised learning for skin lesion segmentation on the ISIC 2018 benchmark dataset. A teacher U-Net with a pretrained ResNet-34 encoder is first trained on a labeled subset and subsequently used to generate high-confidence pseudo-masks for unlabeled images using confidence thresholding. A student model is then trained on the combined labeled and pseudo-labeled data. With only 20% labeled data (440 images), the semi-supervised student achieves a Dice score of 0.7638, recovering 98.3% of the fully supervised baseline (Dice 0.7769). With 50% labeled data (1,102 images), the student achieves a Dice score of 0.7800, demonstrating performance comparable to the fully supervised baseline. A systematic threshold sensitivity analysis across τ ∈ {0.60, 0.70, 0.80, 0.90} identifies τ = 0.80 as the optimal balance between pseudo-label quality and quantity. We release a fully open-source and reproducible implementation that runs on free Kaggle GPU resources, enabling the community to replicate, verify, and extend our findings without specialized hardware.