Abstract
In medical imaging, researchers often require accurate segmentations for specific unlabeled datasets rather than a universal model for future unseen data. This paper investigates transductive learning in this context, specifically focusing on the self-training framework. By analyzing self-training through the lens of Information Gain (IG), the authors show that model calibration is critical: over-confident models inject negative information, while well-calibrated ensembles provide robust pseudo-labels. Experimental results on traumatic brain injury (TBI) lesion segmentation demonstrate that transductive ensemble-based self-training significantly outperforms standard inductive methods.
The Case for Transduction in Medical Imaging
Most semi-supervised learning (SSL) research focuses on induction—training a model on a mix of labeled and unlabeled data to improve performance on future, unseen cases. However, many clinical scenarios, such as large-scale population studies or retrospective phenotype extraction, prioritize the most accurate segmentations for a specific, existing database. This is the domain of transductive learning.
The challenge in transduction is effectively leveraging the unlabeled "target" data during optimization to maximize prediction quality on that exact set. The authors argue that despite its practical utility, transduction is often overlooked in favor of the more general inductive paradigm.
Methods: Information Gain and Ensemble-based Self-Training
The researchers utilize a self-training framework where a model, initially trained on labeled data, generates "pseudo-labels" for unlabeled data to guide its own retraining. They introduce two primary refinements to this process:
- Information Gain (IG) Analysis: The authors provide a theoretical framework demonstrating that self-training is beneficial only when the Information Gain is positive. A model that is over-confident (assigning high certainty to incorrect labels) introduces "negative information," which degrades performance. Conversely, "under-confident" models are safer as they avoid creating confidently wrong pseudo-labels.
- Ensemble-based Self-Training: To mitigate the over-confidence typical of deep neural networks, the authors employ ensembles—multiple models trained with different initializations. Ensembles yield better-calibrated uncertainty estimates.
As shown in Figure 1, the transductive pipeline involves using an ensemble's averaged predictions, filtered by a confidence threshold, to create pseudo-labels for the target data.
Experimental Setup
The study used a multi-center MRI database for multi-class segmentation of TBI lesions:
- DB1: 180 labeled subjects from a single clinic for initial training.
- DB2: 101 unlabeled subjects from 9 clinics used for transductive optimization.
- DB3: 265 subjects from 14 sites used for a final blinded comparison against expert manual refinements.
Results and the Calibration Connection
The results highlight a strong link between model calibration and the success of self-training. While single models are highly sensitive to the chosen confidence threshold ($t$), ensemble-based approaches are remarkably robust.
Key Findings:
- Performance Boost: In the transductive setting, ensemble-based self-training achieved an average Dice score of 53.3%, significantly higher than the supervised baseline (43.3%) and the inductive ensemble (50.0%).
- Calibration Matters: As illustrated in Figure 6(c), ensembles are much better calibrated than single models. This calibration allows them to maintain high performance across a wide range of confidence thresholds (Figure 6(b)), whereas single models fail if the threshold is not perfectly tuned (Figure 6(a)).
- Expert Validation: On the blinded DB3 set, transductive segmentations showed a significantly higher overlap with expert-refined labels (79.4%) compared to inductive ones (76.8%), with $p < 0.05$.
Conclusion
This work demonstrates that for specific clinical research tasks, transductive learning is superior to standard induction. By using ensembles to ensure well-calibrated uncertainty, self-training becomes a robust tool for extracting high-quality labels from unlabeled target data. While the computational cost of ensembles is higher, the gain in accuracy and reliability—especially in correcting systematic errors made by inductive models—makes it a valuable approach for large-scale medical image analysis.
Summary updated