Uncertainty-aware semi-supervised learning for neurosurgical navigation

Abstract

Background/Objective: Accurate, real-time segmentation of anatomical structures during neurosurgical procedures can support intraoperative orientation. One of the most significant challenges in this domain is developing robust segmentation models with limited annotated data while maintaining clinical reliability. This work addresses how semi-supervised learning can leverage both labeled and unlabeled data, while ensuring the dependability crucial for clinical applications where even small segmentation errors can have significant consequences. Methods: We present a novel uncertainty-aware semi-supervised framework for neurosurgical scene segmentation. Our approach introduces Semantic Spatial Uncertainty (SSU), a metric that quantifies prediction reliability by analyzing spatial consistency across multiple stochastic forward passes using Monte Carlo Dropout. The framework employs class-specific calibration with adaptive thresholds that continuously refine through iterative pseudo-labeling, effectively counteracting dataset imbalance. Results: Our method achieves significant improvements for clinically critical classes, with relative gains in Dice Similarity Coefficient of + 40% for tumors, + 15% for middle cerebral artery and + 14% for aneurysm. Unlike traditional uncertainty measures, SSU captures uncertainty even for structures with high perimeter-to-area ratios, demonstrating strong correlation with segmentation quality (Pearson coefficient -0.85) without requiring ground truth. Our approach also outperforms intensive data augmentation (even at 200% synthetic samples) and maintains effectiveness across multiple architectures, demonstrating its architecture-agnostic advantages. Conclusion: By reframing annotation scarcity as an uncertainty quantification problem, our approach provides a practical solution for medical image segmentation in data-constrained environments. This segmentation framework offers potential applications beyond neurosurgery to other computer vision segmentation tasks with limited labeled data. Code is available at htt