Morphological variations and differences of the mandible and midfacial skeleton across different sagittal skeletal patterns and sexes: a three-dimensional geometric morphometric analysis Authors: Guangpu Zhao, Xinyu Su, Danqing He, Yan Gu Published: 2026 ABSTRACT BackgroundMaxillofacial morphological analysis is essential for orthodontic diagnosis, treatment planning, and outcome evaluation. However, previous two-dimensional or landmark-based studies can hardly comprehensively analyze three-dimensional morphology. Therefore, this study aimed to analyze morphological variations and differences of the mandible and midfacial skeleton across different sagittal skeletal patterns and sexes using three-dimensional geometric morphometrics, while also providing novel perspectives and methodologies for morphological analysis in orthodontics.MethodsThe midfacial skeleton and mandible were analyzed separately. Three-dimensional dense templates were first constructed and template registration was validated. Through template registration, all segmented skeleton images (n = 186) obtained a homologous configuration with tens of thousands of quasi-landmarks. After generalized Procrustes analysis, principal component analysis (PCA) was conducted to examine shape variations. The average skeletal shapes were calculated and compared according to sagittal skeletal pattern and sex. Partial least squares regression (PLSR) was used to evaluate the independent effects of ANB angle and sex on skeletal shape.ResultsPCA revealed the principal patterns of shape variations. The first principal component (PC1) of the mandible primarily showed variation in sagittal length and inter-ramal width (explaining 31.80% of the total variance), whereas PC1 of the midfacial skeleton mainly showed variation in vertical height and inter-zygomatic width (24.07%). Morphological differences were demonstrated through comparisons of average skeletal shapes grouped by sagittal skeletal pattern or sex. PLSR indicated significant effects of ANB angle and sex on the three-dimensional morphology of both the mandible and midfacial skeleton (all P < 0.001). Specifically, with increasing ANB angle (skeletal Class II tendency), the length of the condyle (condylar head and neck) decreased, the antegonial notch became more prominent, and the maxillary alveolar bone plane rotated downward and backward (clockwise). Compared with females, males exhibited a more prominent mandibular angle, increased inter-zygomatic width, and more protrusive nasal bridge and anterior nasal spine.ConclusionsThree-dimensional dense geometric morphometrics enables a more comprehensive analysis of facial skeletal morphology. Three-dimensional morphological variations and differences of the mandible and midfacial skeleton across different sagittal skeletal patterns and sexes were characterized.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s40510-026-00645-y. Background The midfacial skeleton and mandible play a crucial role in defining facial morphology. Abnormalities and disharmony in these two structures may compromise dental and facial esthetics, impair oral function, and negatively affect psychological well-being and quality of life [1, 2]. For orthodontic treatment in children, monitoring and controlling maxillofacial growth is very important. Additionally, maxillofacial morphological analysis is essential in orthodontic treatment planning and outcome evaluation. Therefore, a comprehensive understanding of the three-dimensional (3D) morphology of the facial skeleton and associated factors is of considerable clinical importance. The sagittal skeletal pattern describes the sagittally positional relationship of the maxilla and mandible and is widely used in orthodontics and orthognathic surgery for diagnosis and treatment planning. Orthodontists are concerned with skeletal characteristics associated with sagittal skeletal patterns, such as alveolar bone morphology [3], basal and dentoalveolar transverse widths [4], mandibular asymmetry [5], and size [6] and position [7] of condyle. However, previous studies mainly focused on localized features, lacking comprehensive analyses of morphology of midfacial skeleton and mandible. In addition to sagittal skeletal pattern, sexual dimorphism is an important factor influencing the facial skeletal morphology [8]. Understanding the morphological differences of facial skeleton associated with sagittal skeletal pattern and sex can help clinicians identify skeletal abnormalities, guide orthodontic treatment, and inform orthognathic surgery planning and skeletal reconstruction. Moreover, such knowledge can provide important insights into maxillofacial growth patterns. Many studies have analyzed facial skeletal characteristics and their associations with sagittal skeletal patterns and sex, using linear and angular measurements [4–7, 9]. However, these measurements captured only partial aspects of the structures and failed to fully represent the morphology. Geometric morphometrics (GMM) is a method for analyzing overall structural shape based on homologous landmarks and has been widely used to characterize complex biological structures. In orthodontics, it has been used to study craniofacial growth patterns [10, 11], shape changes of hard and soft tissues during orthodontic treatment [12], and characteristics of different sagittal skeletal patterns [13–15]. However, the traditional GMM applied to facial skeleton has two main limitations: First, it relies on two-dimensional images, such as lateral cephalometric radiographs, which are insufficient to represent the 3D morphology and are prone to structural overlap and distortion. Second, it is based on sparse anatomical landmarks, which fail to fully capture the morphological variations of the facial skeleton [16]. Structures such as the condyle or alveolar bone have only a few definable landmarks, making it difficult to characterize their complex morphology. In contrast, 3D dense GMM utilizes 3D images and dense configurations of homologous quasi-landmarks to quantify morphology across the entire anatomical surface, allowing localized and spatially continuous variations between conventional landmarks to be captured. It has been applied to the analysis of facial surface [17, 18] and palatal morphology [19, 20]. However, until now, few studies have explored the application of 3D dense GMM to the midfacial skeleton and mandible. In recent years, the development of the open-source MeshMonk toolbox has enabled the accurate and automatic identification of numerous homologous quasi-landmarks on the face and skeletal surface through template registration [21, 22], which can provide a foundation for 3D dense GMM. This open-source toolbox has been widely adopted in morphological research [23, 24]. It can also be used to establish standardized morphology datasets, which can be used to guide surgical planning [25]. Previously, we have constructed a dense midfacial skeletal template to analyze asymmetry [26]. However, an integrated workflow combining cone beam computed tomography (CBCT)-derived skeletal models, validated dense template registration, and complementary principal component analysis (PCA)- and partial least squares regression (PLSR)-based shape analyses remains underexplored for the characterization of the mandible and midfacial skeleton. This study aimed to use 3D dense geometric morphometrics to comprehensively analyze morphological variations and differences of the mandible and midfacial skeleton across different sagittal skeletal patterns and sexes. Specifically, CBCT-derived models of the two skeletal structures were analyzed separately. Validated dense template registration was used to establish homologous configurations comprising tens of thousands of quasi-landmarks. PCA was used to characterize the principal patterns of shape variation, whereas average-shape comparisons and PLSR were used to characterize morphological differences and quantify the independent associations of ANB angle and sex with skeletal shape, respectively. A standardized facial skeletal morphology dataset and average skeletal shapes were developed to serve as valuable references for identifying skeletal abnormalities. Furthermore, the study sought to provide novel perspectives and methodologies for three-dimensional morphological analysis in the field of orthodontics. Materials and methods The simplified workflow was depicted in Fig. 1A. Fig. 1Partial overview of the methods. A, Simplified workflow of the study. The workflow was applied separately to the midfacial skeleton and the mandible. CBCT: cone beam computed tomography; 3D: three-dimensional; GPA: generalized Procrustes analysis; PCA: principal component analysis; PLSR: partial least squares regression. B, The segmentation range and validation landmarks. Segmentation range: the midfacial skeleton included the maxilla, nasal bone, and zygomatic bone but excluded the teeth; the mandible excluded the teeth. Landmarks: 1, Frontomalare orbitale; 2, Frontomalare temporale; 3, Jugale; 4, Zygotemporale superior; 5, Zygotemporale inferior; 6, Middle zygomaticotemporal suture; 7, Orbitale; 8, Alare; 9, Maxilla point; 10, Nasion; 11, Anterior nasal spine; 12, Subspinale; 13, Mid-philtrume; 14, Prosthion; M1, Condylar superior pole; M2, Condylar medial pole; M3, Condylar lateral pole; M4, Condylar posterior pole; M5, Sigmoid notch deepest point; M6, Coronion; M7, Lingulare; M8, Gonion; M9, Infradentale; M10, Supramentale; M11, Pogonion; M12, Menton; M13, Genion. C, Templates of the midfacial skeleton and mandible and the automatic registration process. After automatic registration, each quasi-landmark corresponded to the same anatomical location across all target skeletons Partial overview of the methods. A, Simplified workflow of the study. The workflow was applied separately to the midfacial skeleton and the mandible. CBCT: cone beam computed tomography; 3D: three-dimensional; GPA: generalized Procrustes analysis; PCA: principal component analysis; PLSR: partial least squares regression. B, The segmentation range and validation landmarks. Segmentation range: the midfacial skeleton included the maxilla, nasal bone, and zygomatic bone but excluded the teeth; the mandible excluded the teeth. Landmarks: 1, Frontomalare orbitale; 2, Frontomalare temporale; 3, Jugale; 4, Zygotemporale superior; 5, Zygotemporale inferior; 6, Middle zygomaticotemporal suture; 7, Orbitale; 8, Alare; 9, Maxilla point; 10, Nasion; 11, Anterior nasal spine; 12, Subspinale; 13, Mid-philtrume; 14, Prosthion; M1, Condylar superior pole; M2, Condylar medial pole; M3, Condylar lateral pole; M4, Condylar posterior pole; M5, Sigmoid notch deepest point; M6, Coronion; M7, Lingulare; M8, Gonion; M9, Infradentale; M10, Supramentale; M11, Pogonion; M12, Menton; M13, Genion. C, Templates of the midfacial skeleton and mandible and the automatic registration process. After automatic registration, each quasi-landmark corresponded to the same anatomical location across all target skeletons Sample patient inclusion The sample was retrospectively collected from patients who visited the Department of Orthodontics, Peking University School and Hospital of Stomatology. The inclusion criteria were as follows: (1) aged ≥ 16 years old; (2) complete permanent dentition and no missing teeth (excluding third molars); (3) available CBCT scans. Patients with congenital craniofacial syndromes, severe facial asymmetry (Menton deviation ≥ 4 mm), severe condylar bone defect, or a history of craniofacial trauma, surgery, or orthodontic treatment were excluded. Finally, 186 patients were enrolled in this study. CBCT scans were obtained using the same CBCT device (Newtom VGi, Quantitative Radiology, Verona, Italy) with parameters set to 110–120 kV, 2–5 mA, voxel size 0.3 mm; and field of view of 16 × 16 cm or 15 × 15 cm. Patients were instructed to sit naturally upright and maintain maximum intercuspation. Data processing Skeletal shapes were segmented from CBCT scans using a two-step strategy, consisting of an initial automated segmentation followed by manual refinement. Automatic segmentation was performed using a previously validated open-source deep learning-based tool, DentalSegmentator, which was recently developed and validated by Dot G et al. [27] After automatic segmentation, an experienced researcher conducted precise manual refinement in Mimics (24.0, Materialise, Leuven, Belgium). The segmentation range of midfacial skeleton included the maxilla, nasal bone, and zygomatic bone, excluding teeth. Mandibular teeth were also removed (Fig. 1B). When necessary, concave defects that did not affect overall shape (e.g. mandibular third molar extraction sockets) were sealed using Geomagic Wrap (2021, 3D Systems, Morrisville, USA). All segmentation results were reviewed by an orthodontic specialist. Finally, 3D skeletal surface images were exported for analysis. Five samples were randomly selected, and manual refinement was repeated by the same researcher after an interval of more than one week. The Dice Similarity Coefficient (DSC) and mean surface distance (MSD) between the two manually refined results was calculated to evaluate intra-operator reliability [28]. Identification and validation of homologous quasi-landmarks in skeletal shapes Geometric morphometrics requires the identification of homologous landmarks and quasi-landmarks across all shapes. In this study, the open-source MeshMonk toolbox was employed to identify dense homologous quasi-landmarks through an initial rigid registration followed by non-rigid registration [21]. In this study, general templates were first constructed, and initial landmarks were placed, followed by automatic registration to identify homologous quasi-landmarks. The accuracy and repeatability of the automatically identified quasi-landmarks were subsequently evaluated. These procedures were based on the methods proposed by White [21] and Wilke [29]. Thirty-one subjects were selected from the total sample for template establishment and validation. The 31 subjects included 13 skeletal Class I, 8 skeletal Class II, and 10 skeletal Class III individuals, comprising 6 males and 25 females, with a mean age of 25.0 ± 6.1 years. The details of template establishment could be found in Supplementary methods [29]. The anterior outer surfaces of the midfacial skeleton that directly support the facial soft tissues were selected as the template range. The anatomical boundaries included the orbital margin, the posterior and inferior margin of zygomatic bone, the zygomaticoalveolar ridge, the inferior margin of alveolar process, and the margin of the piriform aperture [26] (Fig. 1C). For the mandible, the template range was the entire surface after segmentation. Finally, the midfacial skeletal template consisted of 13,482 quasi-landmarks, and the mandibular template consisted of 17,908 quasi-landmarks. The initial landmarks of registration were detailed in Supplementary methods. The templates of midfacial skeleton and mandible were separately registered to the skeleton images of 186 subjects using MeshMonk toolbox in MATLAB (R2023a, MathWorks, Natick, USA), enabling the automatic identification of tens of thousands of homologous quasi-landmarks. Following automatic registration, each quasi-landmark corresponded to the same anatomical location across all target skeletons (Fig. 1C). The validation involved selecting representative landmarks and comparing automatically identified landmarks with those manually identified [21]. 23 landmarks were selected for the midfacial skeleton and 21 for the mandible (Fig. 1B). Landmark definitions were detailed in Supplementary Tables 1 and 2. Each of the two researchers performed landmarking in validation samples for three times. The procedures for manual and automatic placement of validation landmarks were detailed in the Supplementary methods and Supplementary Fig. 1. The Euclidean distances between average manual landmarks (CML) and automatic landmarks (CAU) were calculated to assess the accuracy of the automatic registration. Bland-Altman plots and intraclass correlation coefficients for CML and CAU in x, y, and z axes were also obtained. Root mean square (RMS) distance, calculated as the root square of the mean of squared Euclidean distances of each repeated placement to the centroid of a set of placements [22], was used to assess the reliability. Statistical analysis and shape analysis All statistical and shape analyses were conducted with SPSS Statistics (26.0, IBM, Armonk, USA) and MATLAB (R2023a, MathWorks, Natick, USA). The ANB angle (point A, Nasion, point B) was measured twice with a 48-hour interval by an experienced orthodontic researcher using reconstructed lateral cephalograms in Dolphin 3D software (version 11.95, Dolphin Imaging & Management Solutions, Chatsworth, USA) [30]. A second researcher repeated the ANB angle measurements in 40 subjects. The intraclass correlation coefficient (ICC) for both intra-observer and inter-observer absolute agreement was calculated to evaluate the reliability of the ANB angle. The subjects were divided into three sagittal skeletal patterns based on the average ANB angle: skeletal Class II (ANB > 5°), skeletal Class I (0–5°) and the skeletal Class III (ANB < 0°) [31, 32]. Descriptive statistics were computed for age, sex, and the ANB angle across the three patterns. After identification of homologous quasi-landmarks, shape analyses were conducted separately for the mandible and midfacial skeleton. Due to object symmetry in the mandible and midfacial skeleton, the symmetric component was obtained prior to shape analysis [33]. Since templates with paired quasi-landmarks were used, after template registration, the original configurations were mirrored by reversing the sign of the x-coordinates, and the corresponding paired quasi-landmarks were then relabeled to establish point-to-point homology between the original and mirrored configurations. Generalized Procrustes analysis (GPA) was performed on all original and mirrored configurations to remove differences in position, size, and orientation [34]. The symmetric component was computed as the average of the original and mirrored configurations [33], and symmetric component was used for all subsequent analyses. Therefore, the shape patterns identified in this study represent variations in the symmetric component of the skeletal structures, rather than overall maxillofacial morphology that includes asymmetry. Principal component analysis (PCA) was applied to reduce dimensionality and extract the principal components (PCs) representing the main patterns of shape variation. The first four principal components were visualized. The mean shape was deformed along each PC at -3 and + 3 standard deviations, shown in green and red, respectively. PCA scatter plots were generated to examine the distribution of the sample with different sagittal skeletal patterns and sexes in the shape space. For each of the first four principal components, a two-way ANOVA was performed to assess the effects of sagittal skeletal pattern and sex on the PC scores. Necessary pairwise comparisons among sagittal skeletal patterns were conducted using Tukey’s post-hoc test. Average skeletal shapes were calculated according to sagittal skeletal pattern and sex. Comparisons were conducted between skeletal Class III and Class I, between Class II and Class I, and between males and females. The statistical significance of morphological differences was assessed using permutation tests (n = 10,000) based on the Procrustes distance between group mean shapes [19]. To make the morphological differences clearer, the displacement between the two group mean shapes was amplified three times for visualization. To further investigate the independent effects of the continuous variable ANB angle (°) and the categorical variable sex on overall morphology, partial least squares regression (PLSR) was employed as the regression model. Considering the potential influence of residual growth, age (years) was included in the regression model as a potential confounding factor [17]. The significance of each predictor variable was assessed using permutation tests (n = 10,000) on the partial R2 value [17, 35]. Results were visualized as shape changes along the regression path of each predictor variable, using color-coded maps and deformed shapes. To evaluate localized effects and their statistical significance, partial R² values were also calculated at each quasi-landmark and evaluated using permutation tests, generating corresponding maps of localized partial R² values and statistical significance [17, 35]. Centroid size (CS) was calculated from each unscaled quasi-landmark configuration before GPA. To assess potential allometric effects, the PLSR analyses were repeated with CS included as an additional predictor. To evaluate whether adjustment for CS altered the ANB- and sex-associated shape patterns, cosine similarity was calculated between the corresponding regression vectors obtained from the primary and CS-adjusted models. Because the primary PLSR analyses involved tens of thousands of correlated quasi-landmark coordinates, with effect sizes quantified using partial R² and statistical significance assessed by permutation tests, conventional analytical power calculations were not directly applicable. Therefore, following the rationale proposed by Cardini for sample-size assessment in geometric morphometric studies [36], a sensitivity analysis was performed to evaluate whether the sample size was sufficient to support the above shape analyses. Specifically, 80% of the full sample (n = 148) was randomly selected on three occasions in MATLAB, and all analyses were repeated in each subsample. For each subsample, repeated the same shape analysis procedures, and the results were compared with those obtained from the full sample. The consistency of the regression vector directions and the PC1 directions between the subsamples and the full sample was evaluated using cosine similarity. To evaluate whether the identified shape variations were sensitive to the choice of sagittal indicator, the PLSR analyses were repeated by replacing ANB angle with Beta angle [37]. Since increasing ANB angle and increasing Beta angle represent opposite directions of sagittal skeletal variation, the absolute cosine similarity between the corresponding regression vectors was calculated. Results The demographic characteristics of the subjects were presented in Table 1. The DSC between repeated manual refinements was 95.6 ± 2.1% (mean ± SD) for the midfacial skeleton and 99.5 ± 0.3% for the mandible. The corresponding MSD was 0.15 ± 0.09 mm for the midfacial skeleton and 0.04 ± 0.02 mm for the mandible. The intra-observer and inter-observer ICCs for the ANB angle were 0.99 (P < 0.001) and 0.96 (P < 0.001). Table 1Demographic information of the subjectsVariablesskeletal Class Iskeletal Class IIskeletal Class IIIPatients (n)795849Sex (n), Male/Female23/5619/3917/32Age (year), mean ± SD24.3 ± 5.325.5 ± 6.923.3 ± 5.5ANB angle (°), mean ± SD2.9 ± 1.26.5 ± 1.4-3.3 ± 2.5Note. Chi-square test was used to compare the sex distribution among the three sagittal skeletal patterns. One-way ANOVA was used to compare the age among the three sagittal skeletal patterns. There were no significant differences in sex distribution or age among the three skeletal patterns (P > 0.05) Table 1Demographic information of the subjectsVariablesskeletal Class Iskeletal Class IIskeletal Class IIIPatients (n)795849Sex (n), Male/Female23/5619/3917/32Age (year), mean ± SD24.3 ± 5.325.5 ± 6.923.3 ± 5.5ANB angle (°), mean ± SD2.9 ± 1.26.5 ± 1.4-3.3 ± 2.5Note. Chi-square test was used to compare the sex distribution among the three sagittal skeletal patterns. One-way ANOVA was used to compare the age among the three sagittal skeletal patterns. There were no significant differences in sex distribution or age among the three skeletal patterns (P > 0.05) Demographic information of the subjects Note. Chi-square test was used to compare the sex distribution among the three sagittal skeletal patterns. One-way ANOVA was used to compare the age among the three sagittal skeletal patterns. There were no significant differences in sex distribution or age among the three skeletal patterns (P > 0.05) Accuracy and reliability of automatically identified landmarks The accuracy and reliability of the automatic landmarks of the midfacial skeleton and mandible were presented in Supplementary Tables 3 and 4, respectively. In the validation sample, the mean Euclidean distances between average manual landmarks (CML) and automatic landmarks (CAU) were 0.79 mm for the midfacial skeleton and 1.05 mm for the mandible. Bland-Altman plots showed that most data points fell within the 95% limits of agreement (Supplementary Figs. 2 and 3). In the validation sample, the mean RMS distances of the automatic landmarks were 0.004 mm for the midfacial skeleton and 0.006 mm for the mandible, both substantially smaller than the intra- and inter-observer errors of the manual landmarks. Analysis of mandibular morphology Since GPA was performed, subsequent analyses were based on relative shape, with size differences excluded. The shape variations explained by the first 4 principal components (PCs) of the mandible were illustrated in Fig. 2A. PC1 (+) described decreased mandibular sagittal length, a retrusive chin, and increased inter-ramal width (31.80% of the total variance). PC2 (+) described a larger ramus-body angle (similar to Co-Go-Me) (18.00%). PC3 (+) described increased mandibular body volume, superior projection of the coronoid process and inferior displacement of the condyle (8.75%). PC4 (+) mainly described a prominent mandibular angle and increased width of lower mandibular body (6.50%). Fig. 2Principal component analysis of mandibular shape. A, Shape variations described by the first 4 principal components (PCs). The mean shape was deformed along each PC at -3 and + 3 standard deviations, shown in green and red, respectively. The numbers indicated the percentage of total variance explained by each PC. B, The scatter plots of mandibular sample in shape space. Green points: skeletal Class I; Purple points: skeletal Class II; Orange points: skeletal Class III; Red points: males; Blue points: females. The center of the cross indicated the mean, and the crosses showed the standard deviation Principal component analysis of mandibular shape. A, Shape variations described by the first 4 principal components (PCs). The mean shape was deformed along each PC at -3 and + 3 standard deviations, shown in green and red, respectively. The numbers indicated the percentage of total variance explained by each PC. B, The scatter plots of mandibular sample in shape space. Green points: skeletal Class I; Purple points: skeletal Class II; Orange points: skeletal Class III; Red points: males; Blue points: females. The center of the cross indicated the mean, and the crosses showed the standard deviation The scatter plots of mandibular sample in shape space were shown in Fig. 2B. PC1 scores showed a significant decrease from Class II to Class I to Class III (P < 0.01). PC3 scores were significantly higher in Class II subjects than in Class I and III subjects (P = 0.005 and P < 0.001) (Supplementary Tables 5 and 6). Males exhibited significantly lower PC1 scores (P = 0.030) and significantly higher PC4 scores compared with females (P < 0.001) (Supplementary Table 5). Differences in PC scores reflected morphological differences along the corresponding shape variation pattern. In conjunction with Fig. 2A, morphological characteristics associated with sagittal skeletal pattern and sex could be observed. Compared with the average mandibular shape of skeletal Class I, Class III presented with a smaller inter-ramal width, increased length of condyle (condylar head and neck), and a reduced mandibular body thickness (P < 0.001) (Fig. 3A). Compared with average mandibular shape of skeletal Class I, Class II exhibited a more prominent mandibular angle and antegonial notch, along with decreased length of condyle (P < 0.001) (Fig. 3B). Compared with females, males exhibited a more prominent mandibular angle, a more pronounced antegonial notch, reduced inter-condylar width, and a more prominent temporal crest (crista temporalis). (P < 0.001) (Fig. 3C). In addition, the six average mandibular shapes based on sagittal skeletal pattern and sex were shown in Supplementary Fig. 4. Fig. 3Comparison of average mandibular shapes. To make the differences more apparent, the visualization was performed using three times the differences between the two average shapes. The exaggerated second average shape was generated by adding three times the differences to the first average shape. A, Green: Average mandibular shape of skeletal Class I (n = 79); Orange: Exaggerated average mandibular shape of skeletal Class III (n = 49). B, Green: Average mandibular shape of skeletal Class I (n = 79); Purple: Exaggerated average mandibular shape of skeletal Class II (n = 58). C, Blue: Average mandibular shape of females (n = 127); Red: Exaggerated average mandibular shape of males (n = 59) Comparison of average mandibular shapes. To make the differences more apparent, the visualization was performed using three times the differences between the two average shapes. The exaggerated second average shape was generated by adding three times the differences to the first average shape. A, Green: Average mandibular shape of skeletal Class I (n = 79); Orange: Exaggerated average mandibular shape of skeletal Class III (n = 49). B, Green: Average mandibular shape of skeletal Class I (n = 79); Purple: Exaggerated average mandibular shape of skeletal Class II (n = 58). C, Blue: Average mandibular shape of females (n = 127); Red: Exaggerated average mandibular shape of males (n = 59) PLSR results showed that, after controlling for the effects of sex and age, the ANB angle explained 14.5% of the total shape variations in the mandible (Partial R² = 0.145, P < 0.001). As the ANB angle increased, the chin became more retrusive, the inter-ramal width widened, the posterior border of ramus shortened—particularly the length of condyle—and antegonial notch became more prominent. (Fig. 4A and B). The proportion of local shape variation explained by ANB angle was highest in the condylar process and chin (Fig. 4C), and statistically significant regions were observed throughout the mandibular surface (Fig. 4D). Moreover, after controlling for the effects of ANB angle and age, sex explained 2.1% of the total shape variations in the mandible (Partial R² = 0.021, P < 0.001). The sex-related differences were consistent with the results of the average shape comparisons (Fig. 4E and F). The proportion of local shape variation explained by sex were relatively high in mandibular ramus, particularly in the temporal crest and mandibular angle (Fig. 4G), with statistical significance observed in these regions (Fig. 4H). Fig. 4Independent effects of ANB angle and sex on mandible by partial least squares regression. A, Shape changes (mm) along the surface normal per 1° increase in ANB angle; red indicated outward displacement, blue indicated inward displacement. B, The mean shape was deformed along the ANB angle regression path at -3 and + 3 standard deviations, shown in yellow and purple, respectively. Sample standard deviation of ANB angle was 4.12°. C, Localized partial R² map showing the proportion of shape variation at each quasi-landmark explained by ANB angle. D, Regions showing statistically significant localized effects (P < 0.05, shown in red), as determined by permutation testing with 10,000 permutations. E, Shape changes (mm) along the surface normal from males to females; red indicated outward displacement, blue indicated inward displacement. F, The mean shape was deformed along the sex regression path at -1.5 and + 1.5 units, shown in yellow and purple, respectively; the yellow-to-purple corresponded to triple the male-female differences. G, Localized partial R² map showing the proportion of shape variation at each quasi-landmark explained by sex. H, Regions showing statistically significant localized effects (P < 0.05, shown in red), as determined by permutation testing with 10,000 permutations Independent effects of ANB angle and sex on mandible by partial least squares regression. A, Shape changes (mm) along the surface normal per 1° increase in ANB angle; red indicated outward displacement, blue indicated inward displacement. B, The mean shape was deformed along the ANB angle regression path at -3 and + 3 standard deviations, shown in yellow and purple, respectively. Sample standard deviation of ANB angle was 4.12°. C, Localized partial R² map showing the proportion of shape variation at each quasi-landmark explained by ANB angle. D, Regions showing statistically significant localized effects (P < 0.05, shown in red), as determined by permutation testing with 10,000 permutations. E, Shape changes (mm) along the surface normal from males to females; red indicated outward displacement, blue indicated inward displacement. F, The mean shape was deformed along the sex regression path at -1.5 and + 1.5 units, shown in yellow and purple, respectively; the yellow-to-purple corresponded to triple the male-female differences. G, Localized partial R² map showing the proportion of shape variation at each quasi-landmark explained by sex. H, Regions showing statistically significant localized effects (P < 0.05, shown in red), as determined by permutation testing with 10,000 permutations Analysis of midfacial skeletal morphology The shape variations explained by the first 4 PCs of the midfacial skeleton were illustrated in Fig. 5A. PC1 (+) described decreased vertical dimension and increased inter-zygomatic width (24.07% of the total variance). PC2 (+) mainly described increased sagittal length of the midfacial skeleton (12.90%). PC3 (+) described anterosuperior protrusion of the nasal bridge, infraorbital margin, and anterior nasal spine (9.91%). PC4 (+) mainly described alveolar bone retrusion with enlarged maxillary body and zygomatic volume (8.85%). Fig. 5Principal component analysis of midfacial skeletal shape. A, Shape variations described by the first 4 principal components (PCs). The mean shape was deformed along each PC at -3 and + 3 standard deviations, shown in green and red, respectively. The numbers indicated the percentage of total variance explained by each PC. B, The scatter plots of midfacial skeletal sample in shape space. Green points: skeletal Class I; Purple points: skeletal Class II; Orange points: skeletal Class III; Red points: males; Blue points: females. The center of the cross indicated the mean, and the crosses showed the standard deviation Principal component analysis of midfacial skeletal shape. A, Shape variations described by the first 4 principal components (PCs). The mean shape was deformed along each PC at -3 and + 3 standard deviations, shown in green and red, respectively. The numbers indicated the percentage of total variance explained by each PC. B, The scatter plots of midfacial skeletal sample in shape space. Green points: skeletal Class I; Purple points: skeletal Class II; Orange points: skeletal Class III; Red points: males; Blue points: females. The center of the cross indicated the mean, and the crosses showed the standard deviation The scatter plots of midfacial skeletal sample in shape space were shown in Fig. 5B. Class II subjects showed significantly lower PC1 scores than Class I and Class III subjects (P = 0.003 and P = 0.004) (Supplementary Tables 7 and 8). Males exhibited significantly higher PC3 and PC4 scores than females (P < 0.001 and P = 0.006) (Supplementary Table 7). In conjunction with Fig. 5A, morphological characteristics associated with sagittal skeletal pattern and sex could be observed. Compared with average midfacial skeletal shape of skeletal Class I, the region around point A and the piriform foramen was retruded in skeletal Class III, but the difference was not statistically significant (P = 0.138) (Fig. 6A). Compared with average midfacial skeletal shape of skeletal Class I, the most prominent features of Class II were an increased anterior alveolar bone height and a clockwise rotation of the alveolar bone plane in lateral view (P = 0.002) (Fig. 6B). Compared with females, males exhibited wider inter-zygomatic distance, a more protrusive nasal bridge, and a more protrusive anterior nasal spine (P < 0.001) (Fig. 6C). The six average midfacial skeletal shapes based on sagittal skeletal pattern and sex were shown in Supplementary Fig. 5. Fig. 6Comparison of average midfacial skeletal shapes. To make the differences more apparent, the visualization was performed using three times the differences between the two average shapes. The exaggerated second average shape was generated by adding three times the differences to the first average shape. A, Green: Average midfacial skeletal shape of skeletal Class I (n = 79); Orange: Exaggerated average midfacial skeletal shape of skeletal Class III (n = 49). B, Green: Average midfacial skeletal shape of skeletal Class I (n = 79); Purple: Exaggerated average midfacial skeletal shape of skeletal Class II (n = 58). C, Blue: Average midfacial skeletal shape of females (n = 127); Red: Exaggerated average midfacial skeletal shape of males (n = 59) Comparison of average midfacial skeletal shapes. To make the differences more apparent, the visualization was performed using three times the differences between the two average shapes. The exaggerated second average shape was generated by adding three times the differences to the first average shape. A, Green: Average midfacial skeletal shape of skeletal Class I (n = 79); Orange: Exaggerated average midfacial skeletal shape of skeletal Class III (n = 49). B, Green: Average midfacial skeletal shape of skeletal Class I (n = 79); Purple: Exaggerated average midfacial skeletal shape of skeletal Class II (n = 58). C, Blue: Average midfacial skeletal shape of females (n = 127); Red: Exaggerated average midfacial skeletal shape of males (n = 59) PLSR showed that, after controlling for the effects of sex and age, the ANB angle explained 2.4% of the total shape variations in the midfacial skeleton (Partial R² = 0.024, P < 0.001). With increasing ANB angle, the region around point A and the piriform foramen protruded, the inter-zygomatic and posterior alveolar widths were relatively reduced, and the alveolar bone plane rotated downward and backward (clockwise) (Fig. 7A and B). The proportion of local shape variation explained by ANB angle was highest in the anterior alveolar bone (Fig. 7C), and statistical significance was also observed in the regions exhibiting the aforementioned ANB-associated morphological changes (Fig. 7D). Moreover, after controlling for the effects of ANB angle and age, sex explained 2.8% of the total shape variations in the midfacial skeleton (Partial R² = 0.028, P < 0.001). The sex-related differences were consistent with the results of the average shape comparisons (Fig. 7E and F). The proportion of local shape variation explained by sex was relatively high in the zygomatic regions and nasal bridge, followed by the anterior nasal spine (Fig. 7G), with statistically significant effects observed in these regions (Fig. 7H). Fig. 7Independent effects of ANB angle and sex on midfacial skeleton by partial least squares regression. A, Shape changes (mm) along the surface normal per 1° increase in ANB angle; red indicated outward displacement, blue indicated inward displacement. B, The mean shape was deformed along the ANB angle regression path at -3 and + 3 standard deviations, shown in yellow and purple, respectively. Sample standard deviation of ANB angle was 4.12°. C, Localized partial R² map showing the proportion of shape variation at each quasi-landmark explained by ANB angle. D, Regions showing statistically significant localized effects (P < 0.05, shown in red), as determined by permutation testing with 10,000 permutations. E, Shape changes (mm) along the surface normal from males to females; red indicated outward displacement, blue indicated inward displacement. F, The mean shape was deformed along the sex regression path at -1.5 and + 1.5 units, shown in yellow and purple, respectively; the yellow-to-purple corresponded to triple the male-female differences. G, Localized partial R² map showing the proportion of shape variation at each quasi-landmark explained by sex. H, Regions showing statistically significant localized effects (P < 0.05, shown in red), as determined by permutation testing with 10,000 permutations Independent effects of ANB angle and sex on midfacial skeleton by partial least squares regression. A, Shape changes (mm) along the surface normal per 1° increase in ANB angle; red indicated outward displacement, blue indicated inward displacement. B, The mean shape was deformed along the ANB angle regression path at -3 and + 3 standard deviations, shown in yellow and purple, respectively. Sample standard deviation of ANB angle was 4.12°. C, Localized partial R² map showing the proportion of shape variation at each quasi-landmark explained by ANB angle. D, Regions showing statistically significant localized effects (P < 0.05, shown in red), as determined by permutation testing with 10,000 permutations. E, Shape changes (mm) along the surface normal from males to females; red indicated outward displacement, blue indicated inward displacement. F, The mean shape was deformed along the sex regression path at -1.5 and + 1.5 units, shown in yellow and purple, respectively; the yellow-to-purple corresponded to triple the male-female differences. G, Localized partial R² map showing the proportion of shape variation at each quasi-landmark explained by sex. H, Regions showing statistically significant localized effects (P < 0.05, shown in red), as determined by permutation testing with 10,000 permutations Assessment of allometric effects Additional PLSR analyses with centroid size (CS) included as an additional predictor showed that the ANB-associated shape patterns remained highly stable for both the mandible and midfacial skeleton (cosine similarity = 0.984 and 0.999, respectively). In contrast, the sex-associated mandibular regression vector showed lower consistency after adjustment for CS (cosine similarity = 0.641); its partial R² decreased from 2.1% to 1.0%, and the association was no longer statistically significant (P = 0.074). For the midfacial skeleton, the sex-associated shape pattern was generally preserved (cosine similarity = 0.790) and remained statistically significant (P = 0.001), although its partial R² decreased from 2.8% to 2.0%. CS was significantly associated with the shapes of both the mandible and midfacial skeleton (partial R² = 0.019 and 0.018, respectively) (Supplementary Figs. 6 and 7). Sensitivity analysis The results of the sensitivity analysis were presented in Supplementary Figs. 8, 9, and 10. The subsamples showed good agreement with the full sample in both the statistical metrics of the shape analyses and the directions of PC1 and PLSR regression vectors, supporting the adequacy of the sample size for the shape analyses. The PLSR results of using Beta angle as an alternative sagittal indicator were presented in Supplementary Fig. 11. The mandibular shape patterns associated with ANB and Beta angles were highly consistent (cosine similarity = 0.986), with partial R2 values of 14.5% and 16.0%, respectively (both P < 0.001). The corresponding midfacial skeletal shape patterns were also generally consistent (cosine similarity = 0.900), with partial R2 values of 2.4% for ANB angle (P < 0.001) and 1.7% for Beta angle (P = 0.002). The main localized difference was observed in the nasal bridge region. Discussion Identifying three-dimensional maxillofacial characteristics is essential for orthodontic diagnosis, treatment and outcome evaluation. Compared with traditional two-dimensional or landmark-based analyses, three-dimensional dense geometric morphometrics better detects complex morphological variations. This study creatively applied dense geometric morphometrics to a relatively large sample for analyzing three-dimensional morphological variations of the mandible and midfacial skeleton. We identified the main patterns of shape variation in the mandible and midfacial skeleton and characterized their three-dimensional morphological differences across different sagittal skeletal patterns and sexes. These findings may assist orthodontists in distinguishing abnormal from normal facial skeletal morphology and growth patterns. Additionally, a standardized morphology dataset and average shapes of mandible and midfacial skeleton across different sagittal skeletal patterns and sexes were developed. These can serve as valuable references for identifying skeletal abnormalities, thereby informing orthodontic treatment planning, as well as supporting orthognathic surgery planning and skeletal reconstruction. It should be noted that geometric morphometric analysis removes size differences through GPA; therefore, the present results represent only relative shape variations. In addition, the shape patterns identified in this study specifically represent variation in the symmetric component of skeletal shape and do not include asymmetric variation. PCA revealed the principal patterns of three-dimensional shape variations of the mandible and midfacial skeleton. In the orthodontic population included in this study, the predominant variation pattern (PC1) of mandibular shape primarily reflected changes in sagittal dimension, whereas the PC1 of midfacial skeleton mainly showed variation in vertical dimension. Consistent with this observation, PLSR showed that the ANB angle explained 14.5% of the total shape variations in the mandible, compared with 2.4% in the midfacial skeleton. In addition, the PLSR regression direction of the ANB angle showed a higher degree of agreement with PC1 in the mandible than in the midfacial skeleton. These findings suggest that the ANB angle reflects the principal pattern of overall mandibular shape variations, whereas in the midfacial skeleton it primarily explains more localized shape variations. PLSR showed that sex accounted for approximately 2–3% of the total shape variations in both the mandible and the midfacial skeleton. Yamamoto et al. reported in their study of 60 skeletal Class I patients that principal components related to sex determination accounted for 16% of the total shape variations [38]. In the present study, the PLSR results indicated the proportion of total shape variations directly explained by sex. Therefore, the discrepancy arose from differences in the statistical methods, and the results were not contradictory. Since maxillofacial morphology is influenced by multiple factors and dense geometric morphometrics involve high-dimensional data, the PLSR results show that single factor (ANB angle and sex) explains a relatively small proportion of total shape variations, while still potentially exhibiting significant biological and clinical relevance at the local level. In addition to the partial R² and permutation test results for total shape variations, three complementary localized outputs were provided. The regression effect maps illustrate the direction and relative magnitude of shape changes along the regression path of each factor, with larger deformations indicating greater practical significance or potential clinical relevance. The localized partial R² maps show the proportion of local three-dimensional shape variation independently explained by each factor at every quasi-landmark, with higher values indicating greater potential biological relevance. The localized statistical significance maps identify regions with statistically significant effects, as assessed by permutation tests based on the localized partial R² values. In addition, the PLSR findings were integrated with the results of the comparisons of average skeletal shapes. Morphological features that were consistent across the two analyses were further interpreted were further interpreted in conjunction with previous studies and clinical knowledge to facilitate clinically meaningful interpretations. This study compared the average skeletal shapes based on the categorical variable, sagittal skeletal pattern and employed PLSR to assess the effect of the continuous variable ANB angle on morphology. Based on these results, with increasing ANB angle (skeletal Class II tendency), the mandibular morphology exhibited the following changes: (1) Retrusion of the chin. The contour changes of the anterior border of the mandibular symphysis were most prominently characterized by the retrusion of pogonion. (2) A relative increase in the inter-ramal width. From the horizontal perspective, the mandibular ‘V’ opening angle in Class II patients was larger. (3) A decrease in the length of the condyle. Hasebe et al. reported that the condylar height increased from Class II to Class I to Class III based on landmark measurements on CBCT [6], which was consistent with our findings. We also observed a tendency for the condyles to rotate laterally as the ANB angle increased. The mandibular condyle plays a critical role in craniofacial skeletal growth [39, 40]. The observed condylar morphological features associated with sagittal growth patterns may warrant consideration by orthodontists. With increasing ANB angle (skeletal Class II tendency), the midfacial skeletal morphology exhibited the following changes: (1) Protrusion of the region around point A and the piriform foramen, not limited to the alveolar bone. This sagittal difference in the midfacial skeleton among different sagittal skeletal patterns was relatively minor compared with that in the mandible. No significant differences were observed in the average midfacial skeletal shape between Class I and Class III, which may suggest that excessive mandibular sagittal growth plays a greater role in the skeletal Class III facial pattern than insufficient midfacial skeletal growth. Previous study using linear measurements on lateral cephalometric radiographs also showed that maxillary dimensions (palatal length and maxillary height) exhibited no significant difference among the three sagittal skeletal patterns, whereas mandibular body and entire mandibular length were significantly greater in skeletal Class III patients [41]. (2) Reduced inter-zygomatic and posterior alveolar widths. (3) A downward and backward rotation of alveolar bone plane (clockwise rotation). Li et al.’s cephalometric study reported that the occlusal plane in Class II patients was steeper [42]. Braun and Legan reported the downward and backward rotation of the occlusal plane was associated with sagittal occlusal relationship [43]. Therefore, this clockwise rotation of the alveolar bone may serve as a compensatory mechanism for a retrusive mandibular dentition. The results also showed that the rotation pattern was primarily attributed to increased anterior alveolar bone height. It is possible that this morphological feature may be related to deep overbite or gummy smile; however, this potential association requires further investigation. Based on the results of shape analysis, males, compared to females, exhibit: (1) A more prominent mandibular angle (downward and outward). This finding was supported by studies on Greek [44] and Japanese populations [38]. And the prominent mandibular angle may be associated with masseter muscle attachment and greater masticatory muscle force in males. (2) Deepened antegonial notch. Tunis’s study reported that antegonial notch was more prevalent and pronounced in males [45]. (3) A more protrusive nasal ridge and anterior nasal spine. Similarly, Yamamoto et al. reported that males had a more steeply inclined nasal bone [38]. (4) Larger inter-zygomatic width and shorter zygomatic orbital processes. Although GPA removes differences in absolute scale, it does not eliminate allometric covariation between size and shape. The additional PLSR analyses showed that the ANB-associated shape patterns of both the mandible and midfacial skeleton were relatively robust to adjustment for CS. For the mandible, the ANB-associated regression pattern remained highly consistent, although its partial R² decreased from 14.5% to 10.1%. This indicates that the principal ANB-associated mandibular shape pattern was largely independent of size, although its attenuation suggests that a portion of the association was related to allometric effects. In contrast, the sex-associated regression pattern showed lower consistency after adjustment for CS (cosine similarity = 0.641), and its partial R² decreased from 2.1% to 1.0% and was no longer statistically significant. This attenuation suggests that the sex-associated mandibular shape variation was strongly influenced by allometry. CS independently accounted for 1.9% of mandibular shape variations. The shape changes associated with increasing CS were characterized by relatively reduced inter-condylar width, a more prominent mandibular angle, and a more acute sigmoid notch angle. For the midfacial skeleton, the ANB-associated pattern was almost unchanged after adjustment for CS. The sex-associated pattern also remained generally preserved and statistically significant, although its partial R² decreased from 2.8% to 2.0%, indicating that the association between sex and midfacial shape was only partially attenuated after accounting for size. CS independently accounted for 1.8% of the midfacial skeletal shape variations. The shape changes associated with increasing CS were characterized by relative expansion of the maxillary body and reduced relative heights of the zygomatic bones and nasal bridge. Because this was a cross-sectional study, the observed allometric associations reflect between-individual size-shape covariation and should not be interpreted as within-individual longitudinal growth trajectories. The results showed that the accuracy of the automatic identification of the validation landmarks on the midfacial skeleton and mandible was within 2 mm, which is generally considered the clinically acceptable threshold [46]. Furthermore, the errors between the repeated automatic identifications were minimal. Therefore, the automatic registration used in this study can be reliably applied for quantitative morphological analysis of the midfacial skeleton and mandible. In recent years, the accuracy range of automatic craniofacial landmark identification with artificial intelligence (AI) based on CBCT has been approximately 1.5–3 mm [46]. Since this study first conducted segmentation and some landmarks were located on boundaries, the results cannot be compared with those of the artificial intelligence method. However, the method quantifies the complete shape by identifying tens of thousands of quasi-landmarks. The templates excluded the dentition to eliminate the influence of dental variation. Certain highly variable internal structures, like superior and medial walls of the maxillary sinus, are difficult to segment and minimally affect facial morphology. To ensure accuracy, the midfacial skeletal surface supporting soft tissues was selected as the template range. The final templates are compatible with automatic segmentation, and combining automatic segmentation with manual refinement significantly reduced processing time. Although dentition was excluded to focus on skeletal morphology, this precluded direct assessment of dentoalveolar compensation associated with sagittal skeletal pattern. Therefore, the identified alveolar shape patterns should be interpreted as variations in skeletal alveolar morphology rather than as direct evidence of dental compensation. Future studies integrating dense skeletal and dental templates, potentially through the fusion of CBCT and intraoral scan data, may provide a more comprehensive assessment of skeletal-dentoalveolar relationships. We believe that 3D dense GMM holds promise for identifying additional three-dimensional facial skeletal features relevant to orthodontics, and can also serve as a valuable tool for assessing facial skeletal growth and orthodontic treatment outcomes. The dense average mandibular and midfacial skeletal shapes established according to sagittal skeletal pattern and sex may serve as reference templates for future craniofacial research and clinical applications. With the development of age-specific and longitudinal reference templates, this framework may facilitate the assessment of craniofacial growth and developmental trajectories. These reference templates may also support surgical planning by providing quantitative morphological targets. Furthermore, large-scale reference databases could provide a foundation for automated anomaly detection and AI-assisted craniofacial diagnosis. Nevertheless, these reference templates require validation in larger, multicenter, and more diverse populations before clinical application. The ANB angle was selected since it is widely used to classify sagittal skeletal patterns in both clinical practice and research, thereby facilitating clinical interpretation and comparisons across populations and studies. Nevertheless, its value may be influenced by the anteroposterior position of Nasion and facial rotation. Therefore, the identified shape patterns should not be interpreted as exclusively reflecting pure sagittal jaw discrepancy. The additional Beta angle analysis showed that the principal mandibular and midfacial skeletal shape patterns were generally robust across the two sagittal indicators, although a localized discrepancy was observed in the nasal bridge region. This discrepancy may partly arise because a more posteriorly positioned Nasion can result in a larger ANB angle, whereas Beta angle does not incorporate Nasion. Accordingly, since the primary analyses were based on ANB angle, the corresponding findings should be interpreted specifically as shape variations associated with ANB angle. It should be noted that the sample consisted of single-center orthodontic patients and may not fully represent the general population; therefore, the findings should be interpreted with caution when generalizing to other populations. In addition, although both DSC and MSD indicated good reliability of the repeated manual refinements, they are whole-structure summary measures and may not fully capture localized segmentation errors in anatomically complex regions. Future studies should incorporate region-specific surface assessments. Furthermore, classification based on the ANB angle does not account for the relative contributions of the maxilla and mandible to sagittal patterns. Future studies may benefit from stratifying sagittal skeletal patterns according to their underlying skeletal aetiology and subsequently applying three-dimensional shape analysis to provide deeper insights into both intra- and inter-class morphological characteristics. Conclusions (1) In the orthodontic population, the predominant variation pattern (PC1) of mandibular shape mainly showed variation in sagittal length and inter-ramal width, whereas the PC1 of midfacial skeleton mainly showed variation in vertical height and inter-zygomatic width. (2) The mandible and midfacial skeleton exhibited complex and significant three-dimensional morphological differences across sagittal skeletal patterns. Specifically, with increasing ANB angle, the length of the condyle decreased, the antegonial notch became more prominent, and the maxillary alveolar bone plane rotated downward and backward (clockwise). (3) Compared with females, males exhibited a more prominent mandibular angle, increased inter-zygomatic width, and more protrusive nasal bridge and anterior nasal spine. (4) Three-dimensional dense geometric morphometrics enables a more comprehensive analysis of facial skeletal morphology. A standardized morphology dataset and average skeletal shapes across different sagittal skeletal patterns and sexes were developed as valuable references for identifying skeletal abnormalities. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1