Semi-Supervised Learning Allows for Improved Segmentation With Reduced Annotations of Brain Metastases Using...

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Oncology
• Deep Learning for Segmentation

Background:

• Deep learning for brain metastases segmentation requires extensive expert-annotated data.
• Semi-supervised learning (SSL) offers a method to improve model performance with reduced annotation burden.

Purpose of the Study:

• To evaluate the effectiveness of SSL for segmenting brain metastases.
• To compare SSL methods against supervised baselines.

Main Methods:

• Three SSL techniques (mean teacher, cross-pseudo supervision, interpolation consistency training) were adapted using the U-Net architecture.
• Models were trained and evaluated on labeled and unlabeled brain metastases datasets from multiple institutions, using 5-fold cross-validation.
• Performance was assessed using Dice Similarity Coefficient (DSC), Hausdorff distance, and prediction counts.

Main Results:

• SSL methods consistently outperformed supervised baselines across all test sites, with significant DSC improvements (up to 15.4% ± 1.4% on half-sized datasets).
• SSL achieved comparable or superior results to supervised models trained on twice the labeled data in three out of four datasets.
• Improvements were most pronounced on independent external test sets with limited labeled data.

Conclusions:

• SSL is a viable and effective strategy for enhancing brain metastases segmentation performance.
• This approach enables data-efficient deep learning models, crucial for clinical applications where expert annotations are scarce.
• SSL facilitates robust model performance across institutions with varying clinical protocols and imaging scanners.