Specificity- and Calibration-Aware Breast Ultrasound Segmentation via Entropy-Guided Boundary Supervision

Research Question and Approach

This study addresses two persistent challenges in breast ultrasound (BUS) segmentation: boundary leakage in lesion-containing images and high-confidence false-positive activations in normal images. The authors propose a training-level modification where boundary penalties are scaled by per-pixel predictive entropy. This approach forces the model to focus its gradient updates on boundary regions where it is most uncertain, rather than treating all boundary pixels as equally informative. The framework is integrated into a standard U-Net architecture and supplemented with post-hoc spatial temperature scaling to improve probability calibration at inference time.

The proposed method was evaluated on the BUSI dataset against two baselines: a model without boundary supervision and a model with uniformly weighted boundary binary cross-entropy. While mean Dice scores for lesion-containing images remained statistically indistinguishable across all three methods (approximately 0.76), the entropy-guided approach demonstrated a substantial improvement in specificity. Specifically, false-positive activations on normal images dropped from 14/20 and 19/20 in the baseline models to 5/20 with the proposed method. Additionally, spatial temperature scaling reduced the expected calibration error by more than 50%, resulting in more reliable confidence estimates without altering the final segmentation masks.

In clinical practice, false-positive segmentations increase the burden on radiologists and can lead to unnecessary patient anxiety or follow-up procedures. By shifting the focus of boundary supervision to uncertain regions, this study provides a computationally efficient way to improve the reliability of automated segmentation tools. The findings suggest that loss-function refinements can be highly effective at addressing specific failure modes, providing a complementary strategy to architectural advancements like transformer-based encoders.