Label Uncertainty for Ultrasound Segmentation

Inter-observer variability in medical imaging introduces label uncertainty, particularly in modalities requiring subjective interpretation such as lung ultrasound (LUS). We propose a confidence-aware framework for segmentation that incorporates expert-provided per-pixel confidence scores during both labeling and model training. Instead of treating annotations as absolute ground truth, we capture radiologists' uncertainty as a first-class learning signal to model inherent aleatoric uncertainty. Using a novel confidence-annotated LUS dataset, we evaluate segmentation models trained with varying confidence thresholds and show that incorporating label confidence improves both segmentation and downstream clinical performance - including estimation of the S/F oxygenation ratio and prediction of 30-day readmission. Models trained using a 60% confidence threshold outperform conventional training (0% and 50% thresholds), highlighting that learning from highly confident pixels yields superior clinical outcomes. This work demonstrates that explicit modeling of label confidence enhances both the reliability and interpretability of medical AI systems.