Certainty-Aware Skin Lesion Segmentation with Post-Hoc Reliability Estimation for the Segment Anything Model

Abstract

The Segment Anything Model (SAM) represents a major advance in zero-shot visual segmentation, yet it provides purely deterministic outputs without any measure of prediction reliability, a critical limitation for safety-conscious medical imaging applications. This paper introduces a certainty-aware segmentation framework that augments SAM-based zero-shot inference with principled, post-hoc reliability estimation. Three complementary outputs are introduced: a pixel-wise certainty map that identifies spatially localized regions of ambiguity; a global confidence score that provides a scalar measure of overall segmentation trustworthiness; and a quality-flagging mechanism that enables automated screening of unreliable predictions. The framework requires no modification to SAM's architecture and no additional training data, thereby preserving its zero-shot generalization properties. Evaluation on the ISIC 2018 Task 1 skin lesion segmentation benchmark comprising 2,594 dermoscopic images in a fully zero-shot setting yields a mean Dice Similarity Coefficient of 0.820 pm 0.095 and a mean Intersection-over-Union of 0.750 \pm 0.101. A strong positive correlation (Pearson r = 0.84, p < 0.001, n = 2,594) is observed between certainty scores and segmentation quality. High-quality segmentations (DSC> 0.80) are consistently associated with certainty scores above 80%, while low-quality predictions (DSC< 0.70) yield certainty scores below 50%. Stratified analysis confirms a mean DSC difference of over 0.25 between high- and low-certainty tiers (Wilcoxon p < 0.001, Cohen's d = 2.31). These results demonstrate that the proposed certainty metrics reliably track segmentation accuracy and provide a practical mechanism for risk-aware deployment of foundation models in clinical environments.