Calibration, explainability and spatial uncertainty for YOLO-based detection in panoramic dental radiography

Reliable computer-aided analysis of dental panoramic radiographs requires more than high average precision: predicted probabilities must be well-calibrated, explanations should be stable and clinically meaningful, and spatial uncertainty around detections should be explicit. Using a public panoramic X-ray dataset annotated for four common dental conditions–cavity, filling, implant, and impacted tooth–we train modern YOLOv11 variants and study trust-centred properties alongside standard performance. Our best model reaches strong detection metrics (e.g., F1 ≈ 0.81, mAP@0.5 ≈ 0.812), while uncalibrated probabilities show modest overconfidence (ECE ≈ 0.033 at IoU 0.5) that improves with temperature scaling. We compare per-box visual explanations (Grad-CAM, LIME-WRaP, D-RISE, Adaptive Occlusion), revealing consistent patterns but also case-dependent instability that could mislead end-users if taken at face value. Finally, a lightweight “YOLO-σ” head estimating (σx,σy) provides spatial coverage guarantees with a small mAP trade-off, improving coverage–validity alignment for decision thresholds. Together, these results show that trustworthy deployment in dental workflows is feasible but requires explicit calibration, explanation auditing, and uncertainty-aware reporting rather than accuracy alone.