Research map / Label-efficient and robust learning / Semi-supervised learning
Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Systematic benchmark comparing 6 semi-supervised, 7 self-supervised, and 3 supervised methods on four medical image datasets under realistic validation-set sizes and a fixed wall-clock tuning budget, finding MixMatch the most reliable gain.
Problem
Existing semi- and self-supervised learning benchmarks rarely compare both paradigms together, and often tune hyperparameters on unrealistically large labeled validation sets or not at all, so their rankings may not reflect practical utility for medical image classification with scarce labels.
Approach
Compare semi- and self-supervised methods on equal footing using a unified random-search tuning procedure under a fixed wall-clock budget on one A100, with validation sets no larger than the labeled training set and natural class imbalance preserved.
Key results
- MixMatch gives the most reliable gains, with the median gain over the best labeled-set-only baseline of 4.7 points and worst case of -0.2 points across 4 datasets.
- Hyperparameter tuning on validation sets no larger than the labeled train set is effective; all 16 methods show roughly monotonic test improvement over tuning time.
- Recommended methods gain about +5 balanced accuracy points on TissueMNIST, +10 on PathMNIST, and +5 on AIROGS, while most methods add little on TMED-2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MixMatch with unlabeled data | Medical image classifiers · PathMNIST | balanced accuracy gain over best labeled-set-only baseline: +12.4 | Helped |
| Use MixMatch with unlabeled data | Medical image classifiers · TissueMNIST | balanced accuracy gain over best labeled-set-only baseline: +3.7 | Helped |
| Use MixMatch with unlabeled data | Medical image classifiers · TMED-2 | balanced accuracy gain over best labeled-set-only baseline: -0.2 | No clear effect |
| Use MixMatch with unlabeled data | Medical image classifiers · AIROGS | balanced accuracy gain over best labeled-set-only baseline: +5.7 | Helped |
| Use CoMatch with unlabeled data | ResNet-18 classifier · PathMNIST | balanced accuracy gain over best labeled-set-only baseline: +14.1 | Helped |
| Use self-supervised learning with MOCO | Medical image classifiers · TMED-2 | balanced accuracy gain over best labeled-set-only baseline: -18.8 | Hurt |
| Tune hyperparameters on the target dataset rather than transfer them from TissueMNIST (ablation) | MixMatch · AIROGS | balanced accuracy: 61.75 -> 77.14 | Helped |
| Tune hyperparameters on the target dataset rather than transfer them from CIFAR-10 (ablation) | MixMatch · AIROGS | balanced accuracy: Diverged -> 77.14 | Helped |
| Tune hyperparameters on the target dataset | SSL methods · PathMNIST, TMED-2, and AIROGS | balanced accuracy | Mixed |
| Pretrain on ImageNet instead of training from scratch (ablation) | Top-performing methods · Medical image classification tasks | balanced accuracy: less than 3 points of BA | Helped |
| Use ResNet-50 instead of ResNet-18 (ablation) | AIROGS classifiers · AIROGS | balanced accuracy | No clear effect |
Surprising findings
- Pretraining on ImageNet yields only slight (under 3 points BA) final test gains versus training from scratch for top methods, contrary to large gains reported in prior work.
- Transferring hyperparameters from other datasets is highly variable and MixMatch diverged on AIROGS when transferred, showing the importance of tuning.
Limitations
- Focuses on moderate labeled set sizes (30-1000 per class) and ResNet/WideResNet backbones; ViT-based methods not studied.
- Results not verified across multiple train-validation splits due to resource constraints.
- Only balanced accuracy and sensitivity/specificity metrics analyzed; other clinical metrics not studied.
Experimental setup
Datasets: PathMNIST TissueMNIST TMED-2 AIROGS CIFAR-10
Benchmarks: PathMNIST TissueMNIST TMED-2 AIROGS
Metrics: balanced accuracy AUROC partial AUROC sensitivity at 95% specificity
Baselines: Sup MixUp SupCon PseudoLabel MeanTeacher MixMatch FixMatch FlexMatch CoMatch SimCLR MOCO v2 SwAV
Builds on: MixMatch FixMatch SimCLR USB TorchSSL
Closest papers by meaning
- Self-Paced Contrastive Learning for Semi-supervised Medical Image Segmentation with Meta-labels (NeurIPS 2021 · similarity 0.89)
- Towards Realistic Semi-supervised Medical Image Classification (AAAI 2025 · similarity 0.88)
- Medical Image Segmentation with Minimal Labeling Effort: How Far Can We Push the Limits? (AAAI 2026 · similarity 0.87)
- InterLUDE: Interactions between Labeled and Unlabeled Data to Enhance Semi-Supervised Learning (ICML 2024 · similarity 0.86)
- Semi-supervised TEE Segmentation via Interacting with SAM Equipped with Noise-Resilient Prompting (AAAI 2024 · similarity 0.86)
- CGMatch: A Different Perspective of Semi-supervised Learning (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.