Beyond prompt-only segmentation: Foundation model–guided trust-region label refinement for single-stage bone fracture instance segmentation

Accurate instance-level delineation of bone fractures on radiographs is challenging because fractures are ultra-thin, low-contrast trajectories, and dense pixel-wise annotation is expensive. Naively using foundation-model masks as pseudo-label supervision can further degrade performance by introducing boundary noise and off-target regions. We propose a controlled foundation-model integration strategy for thin-structure learning: expert annotations define a narrow trust region, foundation-model predictions act only as a local proposal signal, an instance-wise intersection-over-union quality gate rejects unreliable cases, and minimal morphological closing repairs small gaps while preserving connectivity. This converts noisy foundation-model outputs into a constrained near-boundary refinement signal, always anchored to and preserving the original expert annotations. We additionally evaluate the Medical Segment Anything Model without task-specific training and introduce a skeleton-aware measure that quantifies centerline coverage while penalizing boundary halo over-segmentation. We provide a reproducible benchmark for multi-region bone-fracture instance segmentation across one-stage, two-stage, and transformer-based methods, targeting clinician-facing overlays for triage and reporting. Across five-fold cross-validation, the proposed ground-truth-anchored refined-mask strategy provides a modest improvement in sample-count-matched ablations, while unconstrained MedSAM mask supervision degrades performance. Together, these findings support the importance of trust-region constrained foundation-model guidance for thin-structure supervision, and our final model achieves a mean average precision of 0.63 at an intersection-over-union threshold of 0.50 on a held-out test set.