Development and validation of a novel craniofacial statistical shape model for the virtual reconstruction of bilateral maxillary defects

Abstract

Bilateral maxillary defects are a challenge for fibula free flap reconstruction (FFFR) surgery due to limitations in virtual surgical planning (VSP) workflows. While meshes of unilateral defects can be mirrored to virtually reconstruct missing anatomy, Brown class c and d defects lack a contralateral reference and associated anatomical landmarks. This often results in poor placement of osteotomized fibula segments. This study was performed to improve the VSP workflow for FFFR using statistical shape modeling (SSM) – a form of unsupervised machine learning – to virtually reconstruct premorbid anatomy in an automated, reproducible, and patient-specific manner. A training set of 112 computed tomography scans was sourced from an imaging database by stratified random sampling. The craniofacial skeletons were segmented, aligned, and processed via principal component analysis. Reconstruction performance was validated on a set of 45 unseen skulls containing various digitally generated defects (Brown class IIa–d). Validation metrics demonstrated promising accuracy: mean 95th percentile Hausdorff distance of 5.47 ± 2.39 mm, mean volumetric Dice coefficient of 48.8 ± 14.5%, compactness of 7.28 × 105 mm2, specificity of 1.18 mm, and generality of 8.12 × 10−6 mm. SSM-guided VSP will allow surgeons to create patient-centric treatment plans, increasing FFFR accuracy, reducing complications, and improving postoperative outcomes.