How accurately and reliably can we integrate digital dental models into 3D stereo-photogrammetric images?

Integrating intraoral scans into a three-dimensional (3D)-stereophotogrammetric facial image allows for the construction of a radiation-free dentofacial model. This study aimed to evaluate accuracy and reliability while positioning digital dental models relative to a patient’s three-dimensional facial mesh by using an intermediate facial scan including a transfer jig. Intra-oral scans, two 3D facial meshes (one at rest and one including a digitally designed and 3D printed transfer jig), and a full field-of-view cone beam computed tomography (CBCT) scan were collected from 27 maxillofacial surgery patients (mean age 27.1 years, standard deviation (SD) 10.9, range 15.8–58.8). The intra-oral scans with their accompanying transfer jig were superimposed onto the facial-scan-with-transfer-jig by matching the jig’s extra-oral portion in Viewbox ( www.dhal.com ). The combined structures were then aligned with the facial-mesh-at-rest using the less deformed upper facial region. The golden standard was constructed by first repositioning the intra-oral scans in the segmented CBCT using artificial intelligence ( https://www.creator.relu.eu ), after which the facial-mesh-at-rest was fitted onto the CBCT’s soft-tissue outline. As a final step, the facial meshes at rest of the ground truth, and the jig-assisted superimposition were superimposed as well. The accuracy of the proposed method could then be determined by measuring the translation and rotation (in Euler angles) required to perfectly fit the intra-oral scans of the jig superimposition onto those of the ground truth. The procedure was repeated three times for all patients. Statistical significance was determined using random permutation tests. Linear mean error values varied from − 0.66 mm along the y-axis (SD 0.86) to 0.22 mm along the z-axis (SD 0.80). The average x-axis shift was − 0.16 mm (SD 0.46). The shifts along the y- and z-axis differed significantly from zero ( p = 0.043 and p < 0.0001, respectively). The y-axis rotational error exhibited the smallest mean error (0.03°, SD 0.65). The mean values for the x and z axis were 0.05° (SD 1.13) and 0.68° (SD 1.07) respectively. Only the rotation around the z-axis differed significantly from zero ( p < 0.0001). Between the three repetitions, significant differences were found only for the y- and z-axis shifts between the first and second repetition ( p = 0.030 and p = 0.017). The method demonstrated sufficiently low error values for clinical applications, although inter-individual variability was relatively large.