Structural prompt-response consistency for label-free failure screening in promptable medical segmentation

Promptable segmentation foundation models can return visually plausible masks even when they fail, especially under cross-site shift. Existing reliability estimators often require labelled calibration data, model modification, or pixel-level uncertainty maps. We propose Confidence-Balanced Structural Prompt-Response Consistency (SPRC-CB), a training-free and label-free score that audits a predicted mask by measuring how its structure changes under small, approximately semantics-preserving prompt perturbations, including scale-relative box jitter and a boundary-near negative point. Component, hole, boundary, area-response and negative-prompt instability cues are fused with the model’s own confidence through a balanced rank rule. Across four endoscopic polyp benchmarks, SPRC-CB raises pooled AUPRC over SAM confidence by 0.12–0.19 across three failure definitions (paired bootstrap, all p ≤ 0.005) while retaining comparable or higher AUROC. Rejecting the highest-risk 10 % of masks recovers 73 % of failures versus 51 % for confidence alone. The advantage also persists under noisy boxes, detector-derived prompts and a supporting MedSAM check, as detailed in Supplementary Material S1.