Introduction
In growing patients, reliable quantification of change requires explicitly stating the reference used for superimposition and interpreting all values as relative changes; however, manual workflows are time-consuming and variable. This study assessed measurement reliability and workflow efficiency for reference-explicit analyses, comparing a fully automated, open-source segmentation and registration workflow against a semiautomatic voxel-based approach for clinically useful 3-dimensional assessments.
Methods
Twenty-two Class II patients with cone-beam computed tomographies at pretreatment (T1) and posttreatment (T2) were analyzed. Automated segmentation and voxel-based superimposition were performed with built-in quantitative analysis. Primary outcomes were skeletal and dental changes relative to cranial base and regional maxilla and mandibular superimposition. Three registration approaches incorporating varying levels of artificial intelligence (AI) involvement—conventional, semiautomated, and fully automated—were compared. Performances were assessed by mixed-effects linear regression models.
Results
Agreement for skeletal and dental measurements was high, with minor differences observed between AI-driven and conventional registration approaches, and all methods showed clinically comparable precision. Most absolute average differences between automated and conventional workflows are under 1.5 mm for linear and 1.5° for angular measurements. Cranial base superimposition differences showed an average difference in T2-T1 changes ranging from–0.3 to 0.7 mm, whereas regional superimposition differences showed an average difference in T2-T1 changes ranging from–1.7 to 1.1 mm.
Conclusions
Automated clinician-verified workflow yields reliable and faster 3-dimensional change measures in growing patients. Interpretation must consider the reference region used for superimposition. AI-driven open-source tools provide a practical quantitative analysis to support diagnosis, timing, and assessment of treatment outcomes.
Highlights
-
•
The study compared conventional and automatic registration approaches.
-
•
All methods showed a similar pattern in cranial base and regional superimpositions.
-
•
Results support the clinical application of automated registration in orthodontics.
Three-dimensional (3D) superimposition enables longitudinal documentation of craniofacial changes in growing patients. In clinical orthodontics, practitioners routinely use cranial base and regional maxillary and mandibular superimpositions to visualize changes between time points and communicate progress with patients and teams. , Previous work in growing patient cohorts has relied on cranial base, maxillary, and mandibular superimpositions to quantify displacement patterns, which underscores the expectation that the chosen reference should be stated explicitly. ,,,,
Traditional 2-dimensional cephalometric superimpositions are inherently subjective and qualitative because of projection and measurement constraints: landmark overlap and ambiguity, magnification and parallax, head-posture variability, and the inability to represent transverse changes or true 3D rotations. Even when standardized, these factors introduce observer dependence and hinder reproducibility. In growing patients, an additional challenge is that the anatomic surfaces used for alignment remodel over time; therefore, all measurements must be interpreted as relative displacement with respect to the chosen reference. ,,
Earlier, 3D studies demonstrated the feasibility and value of voxel-based and surface-based superimpositions for quantifying craniofacial change. However, those methods often required multiple software steps, semimanual segmentation and registration adjustments, and substantial diagnostic time, limiting routine clinical friendliness and scalability. The present work builds on that foundation by testing an open-source, automated segmentation and superimposition workflow with a standardized clinician quantitative analysis to reduce time-on-task and operator dependence, reporting the changes relative to reference regions.
Clinics need a standardized, open-source workflow that (1) makes the reference system (cranial base vs regional) balance ,, explicit, (2) produces reliable relative-displacement ,,,,, measurements with a built-in clinician verification step to confirm registration accuracy, and (3) can be completed in a realistic amount of review time. Importantly, such a workflow is intended for measurement standardization and documentation; it is not designed to separate growth from treatment effects or evaluate the effectiveness of specific protocols. This study assesses whether a fully automated, open-source segmentation + registration , workflow, followed by a mandatory clinician verification step for registration accuracy, can deliver agreement-comparable and efficient reference-explicit 3D displacement measurements in growing patients with Class II malocclusions when compared with a semiautomatic voxel-based approach. These innovations will not only streamline workflows but also allow researchers and clinicians to conduct more rigorous morphologic comparisons and make more accurate clinical diagnoses based on unbiased assessments of treatment outcomes. Before these innovations can be implemented in clinical practice, it is essential to evaluate their performance, which is the primary aim of this study.
Material and methods
This retrospective study was designed to assess the performance of cone-beam computed tomography (CBCT) registration approaches incorporating varying levels of artificial intelligence (AI). , Three different registration approaches used are shown in Figure 1 . This multicenter study performs external validation of previously published segmentation and registration methodologies ( Fig 1 ) using 44 CBCT scans from 22 patients with Class II malocclusion at the pretreatment and posttreatment stages to correct maxillomandibular discrepancies at the University of Minnesota. Institutional review board approval was granted by the University of Michigan (HUM00146139). Patients were included if they had a Class II molar relationship with ANB ≥4.75° and SN-GoGn <37°, a cervical vertebral maturation stage of 2-4, and received comprehensive orthodontic treatment with a cervical pull facebow headgear appliance. Exclusion criteria included crossbite, history of facial trauma or medical conditions that affected growth, orthodontic or orthopedic treatment before the baseline, or patients with the need for surgical/extraction treatment. Patient demographics are shown in Table I . For each patient, de-identified full-face large field-of-view CBCT images were acquired at pretreatment (T1) and posttreatment (T2), spanning the pubertal growth spurt.
Registration approaches– conventional, semi and fully automatic registration approaches. This flowchart highlights the conventional and AI-based methodologies for data requirements and validations. Approach 1 was the conventional registration approach using ITK-SNAP to segment regions of reference, and the voxel-based SlicerCMF CMFReg module was used to reregister the T1 and T2 CBCT images. AI-based approaches incorporate AMASSS (Gillot et al ) and Areg (Anchling et al ) technologies, enabling Approach 3 (fully automatic registration) to automatically define the regions of reference and accomplish whole sample registration. Approach 2 (semiautomatic registration) used the user-defined regions of reference and used AReg for automated whole sample registration. This diagram also details the extensive multicenter ground truth labeling performed with 618 and 465 CBCT scans, respectively, segmentation of 3 reference regions for 135 CBCT scans, internal testing and validation with 54 CBCT scans, and external validation involving 44 CBCT scans from 22 patients before and after treatment, which was accomplished in this study.
Table I
Characteristics of study participants
| Characteristics | Participants (N = 22) |
|---|---|
| Age, y | |
| At T1 | 12.5 ± 1.1 |
| At T2 | 14.8 ± 1.1 |
| Gender, n (%) | |
| Female | 12 (54.5) |
| Male | 10 (45.5) |
| Treatment duration, mo | 27.7 ± 7.3 |
Note. All values are presented as mean ± standard deviation or number (percentage).
Because craniofacial structures remodel during growth, all reported changes are expressed as relative displacement with respect to a defined reference. Moreover, 3D superimposition was performed relative to 3 regions of reference—cranial base, mandible, , and maxilla , —to assess overall facial growth and treatment displacements, bone remodeling relative to the cranial base, and regional changes relative to the maxilla and the mandible. ,, For cranial base superimposition, the anterior cranial base was selected as the reference region, given the study population of growing patients and the relative stability of this area during growth. As described by Cevidanes et al, this region includes the following structures: the anterior wall of the sella, anterior clinoid processes, planum sphenoidale, lesser wings of the sphenoid, superior aspect of the ethmoid and cribriform plate, cortical ridges on the medial and superior surfaces of the orbital roofs, and the inner cortical layer of the frontal bones. For maxilla regional superimposition, the reference region was defined according to Ruellas et al. This region includes the maxillary bone bounded inferiorly by the dentoalveolar processes, superiorly by the plane connecting the right and left orbitale points, laterally by the zygomatic process at the orbitale point, and posteriorly by the plane aligned with the distal surface of the second molars. For mandible regional superimposition, the mandible body mask was used as the reference region, as described by Ruellas et al. The region’s boundaries are as follows: inferiorly along the lower border of the mandible, superiorly below the alveolar bone processes, anteriorly at the mandibular symphysis, and posteriorly up to, but excluding the mandibular rami and condyles. This region excludes the teeth, alveolar bone, rami, and condyles, thereby focusing on the basal bone of the mandible along with the stable structures, such as the symphysis and inner cortical layers, while excluding areas prone to remodeling or growth changes, such as the ascending rami, condylar heads, and alveolar regions supporting the teeth.
The image analysis workflow is summarized in Figure 2 , A – E . For approach 1 (conventional registration approach), the image analysis procedure was described in Bates et al. This approach involves segmentations by user interactive ITK-SNAP (version 4.0.1) and voxel-based SlicerCMF registration tool ( cmf.slicer.org ) via 3D Slicer (version 5.2.2). This registration approach has been used in varied studies. , For approach 2, the semiautomatic approach was based on ITK-SNAP segmentation of the regions of reference and a whole sample automated registration approach using 3D Slicer extension Areg ( https://github.com/lucanchling/AREG ). For approach 3, the fully automatic approach was based on open-source, fully automated, whole sample multianatomic skull structure segmentation and registration via 3D Slicer extensions AMASSS23 ( https://github.com/lucanchling/AMASSS_CBCT ) and AReg21, respectively.
CBCT Image processing workflow in the 3D Slicer software: A, Data anonymization. 3D Slicer module Bach Anonymizer was used to remove all identifiable personal information of all the CBCT images; B, Orientation and approximation. 3D Slicer extension Transforms was used to orientate and approximate CBCT images; C, Segmentation. Segmentation of regions of reference for cranial base, maxilla, and mandible could be accomplished by (1) conventional method– manual ITK-SNAP segmentation or (2) AI-based 3D Slicer extension– AMASSS to automatically segment the region of reference; D, Registration. Registration processes use either the (1) conventional method– voxel-based SlicerCMF registration tool or (2) AI-based whole sample automated registration approach using 3D Slicer extension AReg; E, Landmarks and quantification. The landmarks were placed by a gold standard observer and quantified by 3D Slicer extension AQ3DC.
Dental and skeletal landmarks ( Table II ) were identified by a single calibrated examiner on the scans registered relative to the anterior cranial base, the maxillary regional reference, and the mandibular regional reference. To minimize variability from repeated landmark placement, the transformation matrices obtained in approaches 2 and 3 were applied to propagate the same landmark set across all registrations. In other words, for a given patient and time point, a landmark’s coordinates were fixed; only the registration transform changed. This ensured that any differences in linear or angular measurements between among approaches were attributable to differences in registration, not differences in landmark identification. Linear measurements included right-left, anteroposterior, superior-inferior, and 3D Euclidean distances. Angular measurements included yaw, pitch, and roll. All measurements were computed using the AQ3DC extension for 3D Slicer ( https://github.com/DCBIA-OrthoLab/Q3DCExtension ). ,
Table II
Landmark definition ,
| Landmark | Definition |
|---|---|
| Cranial base | |
| Basion (Ba) | The most posteroinferior point of the anterior margin of the foramen magnum in the midsagittal plane. |
| Sella (S) | The most central point of the sella turcica from superoinferior, anteroposterior, and transversal aspects. |
| Nasion (N) | The most anterosuperior junction of the nasofrontal suture. |
| Maxilla | |
| A-point (A) | Deepest concavity near the transversal midline at the anterior maxilla (between incisors and ANS). |
| Anterior nasal spine (ANS) | The most anterior point of the ANS, centered on the posterosuperior and transversal aspects. |
| Posterior nasal spine (PNS) | The most posterior point of the PNS, centered on the posterosuperior and transversal aspects. |
| Maxillary molar cusp tip (UR6, UL6) | The most occlusal point of the mesiobuccal cusp tip of the maxillary right and left first molars. |
| Maxillary molar root apex (UR6, UL6) | The most apical point of the mesiobuccal root apex of the maxillary right and left first molars. |
| Maxillary incisor incisal edge (U1) | Center (mesiodistal and buccolingual aspects) of the incisal edge of the most proclined maxillary central incisor (if both incisors are similar in inclination, the maxillary right central incisor is used). |
| Maxillary incisor root apex (U1) | The most apical point of the root apex of the most proclined maxillary central incisor (if both incisors are similar in inclination, the maxillary right central incisor is used). |
| Mandible | |
| B-point (B) | Deepest concavity near the transversal midline at the anterior mandible (between incisors and Pog). |
| Pogonion (Pog) | The most anterior point of the mandibular symphysis, centered on the posterosuperior and transversal aspects. |
| Gnathion (Gn) | A point on the anterior mandibular symphysis, centered from the posterosuperior and transversal aspects, and constructed perpendicular to the midway point between Pogonion and Menton. |
| Menton (Me) | The most inferior point of the mandibular symphysis, centered on the posterosuperior and transversal aspects. |
| Gonion (Go) | Two points were placed (right and left Gonion) on the most lateral posterior inferior point at an angle of the mandible, constructed point perpendicular to the bisection of the ramus of the mandible and mandibular plane. Midpoint taken between 2 points. |
| Condylion (Co) | Two points were placed (right and left Condylion) on the most lateral posterior superior point of the head of the condyle. Midpoint taken between 2 points. |
| Mandibular molar cusp tip (LR6, LL6) | The most occlusal point of the mesiobuccal cusp tip of the mandibular right and left first molars. |
| Mandibular molar root apex (LR6, LL6) | The most apical point of the mesial root apex of the mandibular right and left first molars. |
| Mandibular incisor incisal edge (L1) | Center (mesiodistal and buccolingual aspects) of the incisal edge of the most proclined mandibular central incisor (if both incisors are similar in inclination, the mandibular right central incisor is used). |
| Mandibular incisor root apex (L1) | The most apical point of the root apex of the most proclined mandibular central incisor (if both incisors are similar in inclination, the mandibular right central incisor is used). |
Statistical analysis
Descriptive statistics were reported in mean and standard deviation. To compare the differences in measurements with the registration approaches, mixed-effects linear regression models were used. The outcomes were linear and angular measurements for each patient. For each measurement, the model assumes random effects for patient and approaches and includes factors at time points (pretreatment [T1] vs posttreatment [T2]) and approaches (conventional, semiautomatic, and fully automatic registrations). This model allowed for the correlation of measurements within patients and among the same measurements from different approaches. For the 3 approaches, differences were computed between the following: (1) semiautomatic vs conventional; (2) fully automatic vs conventional; and (3) fully automatic vs semiautomatic. For each measurement, the 95% confidence intervals for the differences in measurements are reported, with a confidence interval excluding 0 representing a statistically significant difference. P values are reported with values <0.05 as statistically significant. Analyses were conducted using R software (version 4.4.2, 2024-10-31).
Results
Graphic displays between different registration approaches for cranial base and regional maxillary and mandibular superimpositions are shown for 3 cases in Figures 3-5 . The T2-T1 changes were illustrated by the semitransparent overlay between the T2 and T1 surface models. Comparing the T2-T1 changes, case 1 showed notable horizontal changes, and Case 3 exhibited relatively greater vertical changes, whereas case 2 displayed some of both horizontal and vertical changes. For all 3 types of growth and treatment changes, the different registration approaches showed similar visual T2-T1 change patterns.
Cranial base superimposition by 3 registration approaches for 3 distinct patients with Class II malocclusion ( white , time 1 3D rendering; red , time 2 by conventional registration approach; blue , time 2 by semiautomatic registration approach; green , time 2 by fully automatic registration approach). Also note that for each patient, there is a consistency of registration results across different methodologies.
Maxillary regional superimposition by 3 registration approaches for 3 distinct patients with Class II malocclusion ( white , time 1 3D rendering; red , time 2 by conventional registration approach; blue , time 2 by semiautomatic registration approach; green , time 2 by fully automatic registration approach). Also note that for each patient, there is a consistency of registration results across different methodologies.
Mandibular regional superimposition by 3 registration approaches for 3 distinct patients with Class II malocclusion ( white , time 1 3D rendering; red , time 2 by conventional registration approach; blue , time 2 by semiautomatic registration approach; green , time 2 by fully automatic registration approach). Also note that for each patient, there is a consistency of registration results across different methodologies.
For cranial base superimposition ( Table III ), all the 3 approaches suggested similar T2-T1 changes—clockwise rotation changes for both mandibular plane (average T2-T1 changes for 3 approaches: 0.8°, 0.8°, and 0.5°) and palatal plane (1.9°, 1.8°, and 1.6°), posteroinferior change for A point, and anteroinferior change for both B point and Pog. Conventional and semiautomatic registration approaches showed similar angular and linear measurements. For the difference-in-difference estimates, relative to conventional or semiautomatic registration approaches, the T2-T1 changes estimated using the fully automatic registration approach had a greater difference on the sagittal dimension (anteroposterior) compared with the vertical dimension (superoinferior). Using A point as an example, the average T2-T1 changes on the vertical dimension were similar across 3 approaches, whereas the changes on the sagittal dimension were–1.3 mm,–1.2 mm, and–0.7 mm for conventional, semi, and fully automatic registration approaches, respectively. A similar pattern was observed for the B point and Pog, with T2-T1 changes showing more anterior change using the fully automatic registration approach. However, the absolute differences in T2-T1 changes by different approaches are minor, with average linear differences ranging 0.0-0.7 mm.
Table III
Angular and linear T2-T1 changes and the difference of changes by different registration approaches and cranial base superimposition
| Conventional | Semiautomatic | Fully automatic | Semiautomatic vs conventional (95% CI) | P value | Fully automatic vs conventional (95% CI) | P value | Fully automatic vs semiautomatic (95% CI) | P value | |
|---|---|---|---|---|---|---|---|---|---|
| Angular measurements, degree | |||||||||
| Mandibular plane (Go-Gn) | 0.8 ± 1.5 | 0.8 ± 1.4 | 0.5 ± 1.9 | 0.0 (–0.3, 0.3) | 0.995 | –0.3 (–0.6, 0.0) | 0.079 | –0.3 (–0.6, 0.0) | 0.097 |
| Palatal plane (PNS-ANS) | 1.9 ± 1.9 | 1.8 ± 2.0 | 1.6 ± 2.1 | 0.0 (–0.4, 0.3) | 0.966 | –0.2 (–0.6, 0.1) | 0.239 | –0.2 (–0.6, 0.1) | 0.359 |
| Linear measurement, mm | |||||||||
| A-point change | |||||||||
| Anteroposterior | –1.3 ± 1.4 | –1.2 ± 1.3 | –0.7 ± 1.5 | 0.1 (–0.3, 0.5) | 0.730 | 0.6 (0.2, 1.0) | 0.002 | 0.4 (0.0, 0.8) | 0.027 |
| Superoinferior | –4.6 ± 2.2 | –4.6 ± 2.1 | –4.3 ± 2.2 | 0.0 (–0.3, 0.3) | 0.990 | 0.3 (0.0, 0.6) | 0.029 | 0.3 (0.0, 0.5) | 0.042 |
| 3D | 5.0 ± 2.3 | 4.9 ± 2.3 | 4.7 ± 2.2 | –0.1 (–0.3, 0.1) | 0.658 | –0.3 (–0.5, 0.0) | 0.015 | –0.2 (–0.4, 0.0) | 0.135 |
| B-point change | |||||||||
| Anteroposterior | 0.2 ± 1.9 | 0.3 ± 1.7 | 0.8 ± 2.0 | 0.1 (–0.4, 0.7) | 0.823 | 0.7 (0.1, 1.3) | 0.013 | 0.5 (0.0, 1.1) | 0.066 |
| Superoinferior | –5.4 ± 2.8 | –5.4 ± 2.7 | –5.1 ± 2.9 | 0.0 (–0.2, 0.2) | 0.995 | 0.3 (0.0, 0.5) | 0.025 | 0.3 (0.0, 0.5) | 0.033 |
| 3D | 5.8 ± 2.8 | 5.7 ± 2.7 | 5.8 ± 2.6 | –0.1 (-0.5, 0.3) | 0.851 | 0.0 (–0.4, 0.4) | 1.000 | 0.1 (-0.3, 0.5) | 0.862 |
| Pog change | |||||||||
| Anteroposterior | 0.7 ± 1.7 | 0.8 ± 1.6 | 1.4 ± 2.3 | 0.2 (–0.5, 0.8) | 0.844 | 0.7 (0.1, 1.4) | 0.020 | 0.6 (–0.1, 1.2) | 0.085 |
| Superoinferior | –5.8 ± 2.6 | –5.8 ± 2.6 | –5.5 ± 2.7 | 0.0 (–0.2, 0.3) | 0.995 | 0.3 (0.0, 0.5) | 0.031 | 0.3 (0.0, 0.5) | 0.040 |
| 3D | 6.2 ± 2.5 | 6.2 ± 2.5 | 6.3 ± 2.6 | –0.1 (-0.4, 0.3) | 0.897 | 0.0 (–0.3, 0.4) | 0.941 | 0.1 (–0.2, 0.5) | 0.719 |
Stay updated, free dental videos. Join our Telegram channel
VIDEdental - Online dental courses
