Photorealistic 3D Facial Meshes Obtained by Low-Cost Photogrammetry: Are They Precise Enough to Replace Colorless CT-scan Meshes in Virtual Surgical Planning?

Abstract In Virtual Surgical Planning (VSP) for orthognathic procedures, 3D facial meshes are essential for soft tissue analysis and surgical simulation. Meshes obtained via CT segmentation are the standard, but they have the limitation of not incorporating skin texture. Facial 3D meshes obtained using optical devices overcome this limitation by incorporating skin texture, replacing texture-less CT-scan meshes with photorealistic and more patient-friendly 3D models. Since the high cost of these devices discourages widespread adoption, smartphone-based photogrammetry emerges as an accessible and eligible source of lifelike 3D images. However, photogrammetry accuracy in reproducing morphology similarly to CT-Scans may be a concern. This study evaluated the morphological accuracy of photogrammetry-derived texturized meshes to serve as replacement to texture-less CT-scan meshes for VSP. Photogrammetry meshes were generated from 40 facial photographs captured with a conventional smartphone and processed using proprietary (3DZephyr) and open-source (OrtogOnBlender) software. CloudCompare software was used to calculate surface-to-surface deviations of these meshes against the reference CT-based counterpart, determining tridimensional similarity. Individual and mean heatmaps were obtained. Statistical analyses were performed using the paired Wilcoxon test, with a significance level of 5%. Twenty-seven participants (14 males and 13 females) were included. Meshes processed by both software showed highly reliable accuracy, with mean RMSE (“3D error”) of 0.6734mm (±0.1274) for 3DZephyr and 0.6747mm (±0.1210) for OrtogOnBlender, with no statistically significant difference between them (p = 0.7366). These results support smartphone-based photogrammetry as a reliable low-cost source of 3D images, not only for VSP, but also for other specialties demanding precise tridimensional models.