Prior-guided craniofacial soft-tissue reconstruction from CBCT under acquisition uncertainty via identity-quantized shape priors.

Objective.Cone-beam computed tomography (CBCT) suffers from low soft-tissue contrast and metal artifacts, yielding noisy and incomplete soft-tissue surface observations that limit craniofacial modeling for surgical planning. This study aims to reconstruct identity-preserving facial soft-tissue surfaces from CBCT-derived sparse surface evidence for clinical decision support.Approach.We propose a prior-guided reconstruction framework that introduces identity quantization as ananatomicalshape prior to regularize an inherently underdetermined inference problem. By embedding residual vector quantization within a hierarchical encoder, we learn a discrete identity codebook that improves robustness to acquisition-induced outliers and missing regions while preserving patient-specific anatomical structure. A continuous style branch captures fine-scale details, and the two representations are fused to generate detailed meshes.Main results.Evaluation on 490 subjects, including 50 test cases, shows that our method achieves a 34-point landmark distance of 1.53 mm. Geometric accuracy (GA) is confirmed with an L1 Chamfer distance of 1.13 mm and normal consistency of 0.98. A prospective expert study reveals high clinical acceptance, with GA rated 4.27 ± 0.53 (out of 5).Significance.By incorporating an explicit anatomical prior to regularize reconstruction under acquisition uncertainty, our method improves the clinical usability of CBCT-based soft-tissue surface modeling for orthognathic surgery planning.