Semi‐Supervised Learning Allows for Improved Segmentation With Reduced Annotations of Brain Metastases Using Multicenter <scp>MRI</scp> Data

Jon André Ottesen, Elizabeth Tong, Kyrre E. Emblem, Anna Latysheva, Greg Zaharchuk, Atle Bjørnerud, Endre GrøvikJournal of Magnetic Resonance Imaging 2025 · DOI

TL;DR

This paper demonstrates that semi-supervised learning (SSL) significantly improves the segmentation of brain metastases from MRI scans, achieving better results than traditional supervised methods with substantially less labeled data. This approach holds promise for reducing the annotation burden on experts and enabling more robust AI models across diverse clinical settings.

What Problem Does This Paper Address?

Accurate segmentation of brain metastases from MRI scans is crucial for diagnosis, treatment planning, and monitoring disease progression. However, developing high-performing deep learning models for this task typically requires vast amounts of meticulously annotated data, which is time-consuming and expensive to obtain from expert radiologists. This annotation bottleneck limits the widespread adoption and generalization of these AI tools, especially in resource-constrained environments or when dealing with the variability of imaging protocols across different institutions. This paper investigates whether semi-supervised learning, which leverages both labeled and unlabeled data, can overcome this limitation and lead to improved segmentation performance with reduced annotation effort.

Methods

The researchers adapted three popular semi-supervised learning techniques to the U-Net architecture for brain metastasis segmentation:

  • Mean Teacher (MT): This method enforces consistency by minimizing the difference between the predictions of a "student" model and a "teacher" model, where the teacher's weights are an exponentially moving average of the student's weights. This encourages stable and consistent predictions.
  • Cross-Pseudo Supervision (CPS): In this approach, two models are trained simultaneously. They use each other's predictions on unlabeled data as additional "pseudo-labels" to augment the training signal, alongside the ground truth from labeled data.
  • Interpolation Consistency Training (ICT): ICT ensures consistency by enforcing that predictions for a linear combination of two input images match the prediction of the linear combination of their individual predictions.

These SSL methods were compared against a standard supervised baseline. The study also explored the impact of training with both the full labeled dataset and a "half-sized" labeled dataset, while keeping the amount of unlabeled data constant.

Datasets

The study utilized a multicenter dataset comprising labeled and unlabeled MRI scans from four institutions:

  • Stanford University: Provided both a labeled dataset (156 scans) and a large unlabeled dataset (519 scans).
  • Oslo University Hospital (OUH): Provided a labeled dataset (65 scans).
  • University of California, San Francisco (UCSF): Provided a labeled dataset (324 scans).
  • Yale New Haven Hospital (YNHH): Provided a labeled dataset (200 scans).

All subjects in the study had diagnosed brain metastases. The MRI sequences included T1-weighted pre- and post-contrast, and FLAIR, acquired at 1.5 T and 3 T field strengths. The study did not introduce a new benchmark dataset but rather leveraged existing multicenter data to test the generalizability of SSL.

Key Insights and Ideas

  • SSL Outperforms Supervised Learning: The core finding is that semi-supervised learning methods consistently achieve better segmentation performance (measured by Dice Similarity Coefficient - DSC) than traditional supervised methods, even when trained on significantly less labeled data.
  • Reduced Annotation Burden is Achievable: The study demonstrates that SSL can achieve comparable or even better results than supervised models trained on twice the amount of labeled data, highlighting its potential to drastically reduce the need for expert annotations.
  • Generalizability Across Institutions: The SSL models showed improved performance on independent external test sets from different institutions, suggesting that these methods can lead to more robust models that generalize better across varying imaging protocols and scanner types.
  • Impact of Data Size: The benefits of SSL were particularly pronounced when trained on smaller amounts of labeled data. As the amount of labeled data increased, the relative improvement of SSL over the supervised baseline tended to decrease, though it still often remained superior.
  • False Positive Reduction: Notably, when trained with half the labeled data, SSL methods (especially ICT and CPS) showed a reduction in false positives, indicating improved specificity without sacrificing sensitivity.

Key Results

  • DSC Improvement: The best-performing SSL method (CPS) achieved an average DSC improvement of 6.3% to 15.4% when trained on half the labeled dataset across different cohorts, compared to the supervised baseline. When trained on the full dataset, the improvement ranged from 1.8% to 9.2%.
  • Comparable to More Data: In three out of four datasets, SSL training with half the labeled data produced results equal to or better than supervised models trained on twice the labeled data.
  • 95th Hausdorff Distance (HD95): SSL methods also showed a decrease in HD95, indicating more accurate boundary localization, with reductions ranging from -1.5% to -33.5% when trained on half the labeled data.
  • True Positives vs. False Positives: Semi-supervised models generally achieved higher true positives with only a marginal increase in false positives, suggesting an improvement in sensitivity without a significant trade-off in specificity.

Limitations and Future Work

The authors acknowledge several limitations:

  • Limited Model Testing: The study tested only three common SSL methods. Further research could explore other SSL techniques that might offer even greater improvements.
  • Sample Size Impact: The scarcity of labeled data limited the ability to fully examine how varying sample sizes of labeled data influence the observed improvements.
  • P-value Interpretation: The statistical significance (P-values) for DSC was averaged across folds, and more extensive testing with more folds could further validate the findings.

Future work could involve more extensive testing of SSL methods, a deeper investigation into the impact of labeled data sample size, and exploring the robustness of these models to even greater variations in imaging protocols and clinical data. The authors also suggest that the improved accuracy and robustness of SSL models could pave the way for automated annotation and reduced clinical burden, especially in resource-constrained settings.

Summary updated

Abstract

3 TECHNICAL EFFICACY: Stage 2.
BibTeX
@article{ottesen2025semisupervised,
  author = {Jon André Ottesen and Elizabeth Tong and Kyrre E. Emblem and Anna Latysheva and Greg Zaharchuk and Atle Bjørnerud and Endre Grøvik},
  title = {{Semi‐Supervised Learning Allows for Improved Segmentation With Reduced Annotations of Brain Metastases Using Multicenter <scp>MRI</scp> Data}},
  journal = {Journal of Magnetic Resonance Imaging},
  year = {2025},
  doi = {10.1002/jmri.29686},
}
Figure