Source-linked AI summary
Quality-Guided Semi-Supervised Learning for Medical Image Segmentation
Kumar Abhishek, Ghassan Hamarneh
TL;DR
Medical image segmentation is limited by costly dense annotations, while confidence-based SSL reliability estimates are self-referential. The paper trains a quality predictor from image-mask pairs and integrates it through regularization and pseudolabel reweighting, achieving consistent improvements across datasets and architectures.
Problem
Dense medical-image annotations are costly and scarce, while SSL reliability measures based on model confidence lack explicit grounding in segmentation quality.
Method
A dedicated predictor estimates segmentation quality from image-mask pairs and guides SSL through quality regularization and pseudolabel sample reweighting.
Results
The framework consistently improves existing SSL paradigms across multiple datasets, architectures, and settings.
Takeaways & Limitations
Learned, contextually grounded quality prediction provides an effective training signal for leveraging unlabeled medical-image data without modifying the segmentation network architecture.
Takeaways & Limitations
The experiments do not assume matched distributions between labeled and unlabeled data, using related but distinct data sources.
Abstract
from arXiv · showhide
Training accurate medical image segmentation models requires large amounts of densely annotated data, which is costly and time-consuming to obtain. Semi-supervised learning (SSL) alleviates this by learning from both abundant unlabeled data and limited labeled data. However, most modern SSL methods rely on pseudolabels for unlabeled data, and typically assess their reliability through model confidence or uncertainty, measures that are self-referential and lack explicit grounding in segmentation quality. Instead, we propose a quality-guided SSL framework that trains a dedicated network to estimate segmentation quality from image-mask pairs. The predictor is trained on variable-quality masks generated through synthetic corruptions augmented with imperfect outputs from partially trained segmentation models, capturing realistic error patterns encountered during training. We integrate the quality predictor into SSL through two complementary mechanisms: a quality-aware regularization loss and a quality-based pseudolabel sample reweighting scheme. We show that our method serves as a drop-in enhancement to existing SSL frameworks. Extensive experiments across five datasets and multiple architectures demonstrate consistent improvements over competing SSL methods, advancing the state-of-the-art in semi-supervised medical image segmentation.
1 Introduction
Medical image segmentation needs costly dense annotations, motivating SSL with unlabeled data, but confidence-based reliability measures remain self-referential. The paper proposes independently predicted segmentation quality as a grounded learning signal for SSL.
- Dense pixelwise annotations are costly and scarce, motivating semi-supervised learning with limited labels and abundant unlabeled data.
- Existing SSL methods use consistency, pseudolabel, or contrastive objectives, while reliability is often treated uniformly or estimated from model confidence.
- Model-derived confidence remains self-referential and may not detect systematic errors arising from the representations that produced predictions.
- The framework trains a quality predictor on image-mask pairs to provide an independent, image-grounded signal for unlabeled data.
- The method combines a differentiable quality regularizer, pseudolabel reweighting, realistic mask corruptions, and framework-agnostic integration across SSL settings.
2 Method
The method trains a quality predictor on labeled images and variable-quality masks, then freezes it to guide semi-supervised segmentation on labeled and unlabeled data. Quality guidance is integrated through differentiable regularization or pseudolabel reweighting, without architectural changes.
- 2 Method: Training proceeds in two phases: learn gϕ to estimate quality from image-mask pairs, then freeze gϕ while guiding semi-supervised segmentation training.The segmentation model uses both the small labeled set and the larger unlabeled set.
- 2.2 Variable Quality Mask Generation: The quality-predictor training set pairs labeled images with masks degraded by stochastic morphological, geometric, noise, and boundary perturbations.Each corrupted mask receives a Dice score as its quality target.
- 2.2 Variable Quality Mask Generation: Weak segmentation models augment synthetic corruptions with early-training predictions that reproduce characteristic neural-network error patterns.This strategy addresses the limited realism of morphological corruptions alone.
- 2.3 Segmentation Quality Predictor: The predictor compares mask structure with image evidence and provides differentiable quality estimates without requiring ground truth during later use.This contextual grounding distinguishes it from confidence signals based on the segmentation network’s internal state.
- 2.4 Quality-Guided Semi-Supervised Training: For unlabeled samples, QAR propagates quality-loss gradients through predicted masks, encouraging segmentations that the predictor judges high quality.PL-QW instead uses the frozen predictor to compute per-sample weights, upweighting high-quality pseudolabels and downweighting unreliable ones.
- 2.4 Quality-Guided Semi-Supervised Training: Quality-weighted pseudolabeling is orthogonal to the underlying SSL method and can augment pseudolabel approaches by weighting per-sample losses without architectural changes.This makes the mechanism a drop-in enhancement for existing SSL frameworks.
3 Results and Discussion
Across five datasets and three segmentation architectures, quality-guided SSL consistently improves segmentation performance, while quality prediction correlates strongly with measured segmentation quality and benefits from realistic corruptions.
- Overall performance: Quality-guided methods improve segmentation performance across five datasets and three segmentation model architectures.Table 1 compares DSC and IoU for SSL baselines, quality-weighted variants, and QAR.
- Quality prediction: The quality predictor achieves MAE in [0.043, 0.088] and Pearson’s correlation coefficient ρ > 0.92 across all five datasets.These results are reported on the test sets.
- Quality prediction: On CLI, predicted quality and segmentation Dice show a strong positive correlation of ρ=0.69 with p=1e-314.The figure also illustrates successful predictions for near-perfect and poor segmentations, alongside two failure cases.
- Overall performance: QAR outperforms all competing SSL paradigms across the evaluated datasets and models.The comparison includes pseudolabel, mean-teacher, uncertainty-aware, interpolation-consistency, contrastive, and cross-pseudo-supervision methods.
- Quality weighting: Quality-weighted variants consistently outperform their original methods across all but one setting, supporting broad applicability of sample reweighting.The exception is DMF + UN-P.
- Ablations: Weak-model corruptions improve quality prediction, while increasing corrupted masks per sample helps until a saturation point.The ablations also evaluate backbone choice and the effects of quality-guidance weights.
- Ablations: Increasing unlabeled data substantially accelerates convergence but has minimal impact on final DSC.The number of epochs needed to reach 96% validation DSC decreases considerably as M increases.
4 Conclusion
The paper presents a contextually grounded quality predictor for medical image segmentation and integrates it into SSL through regularization and sample reweighting. Experiments across datasets and architectures show consistent improvements over existing SSL paradigms, while future work includes multiclass segmentation and active learning.
- Conclusion: The approach estimates medical image segmentation quality from corrupted masks generated by synthetic degradations and weak segmentation models.The corruption strategy is intended to capture characteristic neural-network errors.
- Conclusion: The quality predictor guides existing SSL frameworks through a regularization loss or sample reweighting without architectural modifications to the segmentation network.These are the two complementary integration mechanisms described in the conclusion.
- Conclusion: Experiments across multiple datasets and model architectures demonstrate consistent improvements over existing SSL paradigms.The authors conclude that learned quality prediction provides an effective training signal for leveraging unlabeled data.
- Future work: Future work could extend quality-guided SSL to multiclass segmentation and use quality predictions to prioritize unlabeled samples for expert annotation.These directions are explicitly identified by the authors.