UNCERTAFACE: A Study Protocol for Evaluating Uncertainty Calibration in Facial Soft-Tissue Reconstruction from Incomplete Cranial Geometry

Reconstructing a face from a skull is an ill-posed problem: the skull does not uniquely determine the overlying soft tissue, and damaged or missing cranial bone removes further evidence. Probabilistic face–skull models and learning-based reconstruction methods already exist, and uncertainty estimation has been studied for cranial bone reconstruction. On a preliminary, non-systematic literature search, and to the best of the author's knowledge, what has not been tested is whether region-specific uncertainty estimates for the reconstructed facial soft-tissue surface are empirically calibrated when the cranium is incomplete, and how that calibration behaves when the pattern of damage at test time differs from the pattern used for calibration. This protocol specifies a study to address that question using paired skull–face geometry derived from CT, with simulated, anatomically structured damage applied to complete skulls; it does not use naturally damaged remains. It defines damage conditions, subject-level data partitioning, regional error and uncertainty measures in millimetres, a split-conformal calibration procedure with per-region marginal coverage targets for upper bounds on regional surface error (not point-wise spatial confidence regions), a damage-pattern-shift stress test, baselines, metrics and statistical analysis. The protocol does not claim that probabilistic reconstruction, uncertainty estimation or conformal prediction is new. Its intended contribution is a pre-specified empirical evaluation in which calibration failures are reported as readily as successes.