Random expert sampling for deep learning segmentation of acute ischemic stroke on non-contrast CT
Abstract
Accurate segmentation of acute ischemic stroke lesions on non-contrast CT scans is crucial for clinical decision-making and treatment planning. However, training deep learning models for this task is challenging due to the subtle appearance of early stroke signs and the need for expert annotations. Traditional approaches often rely on consensus annotations from multiple experts, which can be time-consuming and may not capture the full range of expert interpretations.
We propose a random expert sampling strategy for training deep learning models for acute ischemic stroke segmentation. Instead of requiring consensus among multiple experts, our approach randomly samples from individual expert annotations during training. This strategy exposes the model to diverse expert interpretations and improves robustness to inter-observer variability.
Our experiments demonstrate that random expert sampling leads to improved segmentation performance compared to consensus-based training approaches. The method achieves better generalization across different imaging protocols and patient populations, while maintaining computational efficiency. These results suggest that embracing expert disagreement, rather than forcing consensus, can improve model performance for challenging medical imaging tasks.