Teacher–Student Pseudo-Labeling for Semi-Supervised Skin Lesion Segmentation

This repository contains the pretrained model checkpoints accompanying the paper:

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

The models implement a teacher–student pseudo-labeling framework for semantic segmentation of dermoscopic skin lesions using a U-Net architecture with a pretrained ResNet-34 encoder.

Overview

Pixel-wise annotation of medical images is expensive and requires expert dermatologists. This work investigates a simple yet effective semi-supervised learning strategy that leverages unlabeled dermoscopic images through teacher-generated pseudo labels.

The repository provides pretrained checkpoints for:

Fully supervised baseline
Teacher model (20% labeled data)
Student model (20% labeled + pseudo labels)
Teacher model (50% labeled data)
Student model (50% labeled + pseudo labels)

The implementation is fully reproducible and was trained using free Kaggle Tesla T4 GPU resources.

Model Architecture

Architecture: U-Net
Encoder: ResNet-34 (ImageNet pretrained)
Framework: PyTorch
Image Size: 256 × 256
Task: Binary semantic segmentation
Dataset: ISIC 2018 Skin Lesion Segmentation Challenge

Available Checkpoints

Model	Description
`best_full_supervised_model.pth`	Fully supervised baseline trained using 100% labeled images
`teacher_20pct.pth`	Teacher model trained using only 20% labeled images
`student_20pct.pth`	Student model trained using 20% labeled images and pseudo labels
`teacher_50pct.pth`	Teacher model trained using only 50% labeled images
`student_50pct.pth`	Student model trained using 50% labeled images and pseudo labels

Performance

Fully Supervised Baseline

Metric	Score
Dice	0.7769
IoU	0.6482
Pixel Accuracy	0.9130

Semi-Supervised Results

Experiment	Best Validation Dice
Teacher (20%)	0.7570
Student (20%)	0.7638
Teacher (50%)	0.7759
Student (50%)	0.7800

Using only 20% labeled data, the student model recovers 98.3% of the fully supervised Dice score.

Using 50% labeled data, the student slightly surpasses the fully supervised baseline.

Ablation Study

Confidence threshold analysis for pseudo-label selection:

Threshold	Accepted Masks	Dice
0.60	1743	0.7506
0.70	1685	0.7551
0.80	1563	0.7638
0.90	1154	0.7604

The optimal threshold was τ = 0.80, balancing pseudo-label quality and quantity.

Loading a Checkpoint

import torch

model = YourUNetModel()

checkpoint = torch.load("student_50pct.pth", map_location="cpu")

model.load_state_dict(checkpoint)

model.eval()

Intended Use

These checkpoints are intended for:

Medical image segmentation research
Semi-supervised learning research
Benchmark comparisons
Educational purposes
Reproducibility of the accompanying paper

They are not intended for clinical diagnosis or direct patient care.

Related Resources

📄 Paper (Zenodo): https://doi.org/10.5281/zenodo.21320667
💻 GitHub Repository: https://github.com/HabibaSajid321/semi-supervised-skin-lesion-segmentation
🤗 Model Checkpoints: https://huggingface.co/DevHabiba/skin-lesion-segmentation-unet
🌐 Interactive Demo: https://huggingface.co/spaces/DevHabiba/skin-lesion-segmentation

Citation

If you use this work in your research, please cite:

@misc{sajid2026teacher,
  author       = {Habiba Sajid},
  title        = {Teacher--Student Pseudo-Labeling for Semi-Supervised Skin Lesion Segmentation: A Reproducible Baseline},
  year         = {2026},
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.21320667},
  url          = {https://doi.org/10.5281/zenodo.21320667}
}

---

# License

Released under the MIT License.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

DevHabiba
/

skin-lesion-segmentation-unet