Research map / Label-efficient and robust learning / Semi-supervised learning
Towards Realistic Semi-supervised 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.
Proposes a self-recalibrated semantic training framework for realistic semi-supervised medical image classification with open-set and class-imbalanced unlabeled data, combining informative open-set sample selection, compact semantic clustering, and a dual-balanced classifier.
Problem
Semi-supervised medical image classification in realistic settings where unlabeled data contains unknown (open-set) classes and both labeled and unlabeled data are class-imbalanced, with small inter-class lesion differences.
Approach
Instead of discarding open-set unlabeled samples, select informative ones that share disease-related features with known classes (via prototype-based semantic richness score) and use them as augmentations; add a compact semantic clustering loss and a dual-balanced classifier (similarity pseudo-label regularization plus category-customized thresholds) to handle open-set and imbalance jointly.
Key results
- On ISIC-2019 (γ=5%), the proposed method reaches 69.19% accuracy versus 67.10% for the best closed-set baseline FlexMatch.
- On APTOS-2019 (γ=20%), the method reaches 62.50% vs 61.00% for FlexMatch.
- On HyperKvasir (γ=1%), the method reaches 89.59% vs 88.21% for FlexMatch; ablation shows all three components together give the best results on each dataset.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the full method rather than FlexMatch | Realistic semi-supervised medical image classification · ISIC-2019, γ=5% | accuracy: 67.10±0.54 -> 69.19±0.34 | Helped |
| Use the full method rather than FlexMatch | Realistic semi-supervised medical image classification · ISIC-2019, γ=10% | accuracy: 70.83±0.12 -> 72.27±0.78 | Helped |
| Use the full method rather than FlexMatch | Realistic semi-supervised medical image classification · ISIC-2019, γ=20% | accuracy: 73.53±0.82 -> 74.90±0.70 | Helped |
| Use the full method rather than FlexMatch | Realistic semi-supervised medical image classification · APTOS-2019, γ=5% | accuracy: 53.67±0.85 -> 56.83±2.95 | Helped |
| Use the full method rather than FlexMatch | Realistic semi-supervised medical image classification · APTOS-2019, γ=10% | accuracy: 57.17±1.70 -> 59.00±1.47 | Helped |
| Use the full method rather than FlexMatch | Realistic semi-supervised medical image classification · APTOS-2019, γ=20% | accuracy: 61.00±0.41 -> 62.50±1.22 | Helped |
| Use the full method rather than FlexMatch | Realistic semi-supervised medical image classification · HyperKvasir, γ=1% | accuracy: 88.21±1.19 -> 89.59±1.29 | Helped |
| Use the full method rather than FixMatch | Realistic semi-supervised medical image classification · HyperKvasir, γ=2% | accuracy: 94.31±0.21 -> 94.41±0.82 | Helped |
| Add informative sample selection (ablation) | Proposed framework · ISIC-2019, APTOS-2019, and HyperKvasir | accuracy: 64.7 -> 67.8; 53.0 -> 54.5; 87.0 -> 87.3 | Helped |
| Add compact semantic clustering alone (ablation) | Proposed framework · ISIC-2019, APTOS-2019, and HyperKvasir | accuracy: 64.7 -> 65.3; 53.0 -> 54.5; 87.0 -> 87.5 | Helped |
| Add dual-balanced classifier alone (ablation) | Proposed framework · ISIC-2019, APTOS-2019, and HyperKvasir | accuracy: 64.7 -> 66.4; 53.0 -> 54.0; 87.0 -> 87.2 | Helped |
| Add informative sample selection and compact semantic clustering (ablation) | Proposed framework · ISIC-2019, APTOS-2019, and HyperKvasir | accuracy: 64.7 -> 67.9; 53.0 -> 55.0; 87.0 -> 87.9 | Helped |
Surprising findings
- Open-set samples in unlabeled data reduce performance for several SSL methods (e.g., Pi-Model drops from 65.8 to 60.8 on ISIC-2019 in open-set vs closed-set setting comparison), and indiscriminately using all open-set samples gives only marginal gains.
Limitations
- Evaluation limited to three medical datasets and a single ResNet-50 backbone.
- Open-set classes in the benchmark are constructed by splitting existing datasets, which may not capture all realistic distribution shifts.
- Improvements over the best baseline are modest (about 1-2 points) with overlapping standard deviations in several settings.
Experimental setup
Datasets: ISIC-2019 APTOS-2019 iAMD-Challenge HyperKvasir
Benchmarks: ISIC-2019 APTOS-2019 HyperKvasir
Metrics: accuracy
Baselines: Pi-Model Mean Teacher MixMatch FixMatch FlexMatch CCSSL OpenMatch MTCF T2T
Builds on: FixMatch OpenMatch FlexMatch CCSSL
Closest papers by meaning
- Systematic Comparison of Semi-supervised and Self-supervised Learning for Medical Image Classification (CVPR 2024 · similarity 0.88)
- ACPL: Anti-Curriculum Pseudo-Labelling for Semi-Supervised Medical Image Classification (CVPR 2022 · similarity 0.88)
- A Semantic Knowledge Complementarity based Decoupling Framework for Semi-supervised Class-imbalanced Medical Image Segmentation (CVPR 2025 · similarity 0.87)
- DuSSS: Dual Semantic Similarity-Supervised Vision-Language Model for Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.85)
- Combinatorial CNN-Transformer Learning with Manifold Constraints for Semi-supervised Medical Image Segmentation (AAAI 2024 · similarity 0.84)
- SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.84)
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.