Active Learning in Brain Tumor Segmentation with Uncertainty Sampling and Annotation Redundancy Restriction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Deep learning models show promise in medical imaging analysis.
• High annotation costs and large data volumes limit their clinical application.
• Brain tumor segmentation is a critical but data-intensive task.

Purpose of the Study:

• To compare active learning strategies for minimizing data requirements in 3D brain tumor segmentation.
• To identify the most effective strategy for reducing annotation burden without performance loss.
• To propose a framework for efficient data utilization in medical image analysis.

Main Methods:

• Trained 3D U-net models on 638 multi-institutional brain tumor MRI scans.
• Compared active learning strategies: Bayesian estimation with dropout, bootstrapping, margin sampling, and random query.
• Evaluated annotation redundancy restriction techniques.
• Determined minimum data for performance equivalent to full dataset training (α=0.05).

Main Results:

• Bayesian approximation with dropout achieved target performance using ~30% of the data required by random query (p=0.018).
• Annotation redundancy restriction reduced data needs by 20% compared to random query.
• Dropout uncertainty estimation required the least annotated data for optimal segmentation.

Conclusions:

• Active learning, especially dropout-based uncertainty estimation, substantially reduces annotation burden for 3D brain tumor segmentation.
• This approach enables efficient training of deep learning models with limited data.
• The findings support the clinical translation of AI in medical imaging by addressing data scarcity.