CBCT-Based Morphometric Clustering for Orthodontic Diagnostics Using PCA and K-Means: An Explainable AI Workflow

doi:10.21203/rs.3.rs-8049618/v1

Download PDF

Research Article

CBCT-Based Morphometric Clustering for Orthodontic Diagnostics Using PCA and K-Means: An Explainable AI Workflow

https://doi.org/10.21203/rs.3.rs-8049618/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background: Cone-beam computed tomography (CBCT) enables three-dimensional assessment of craniofacial structures; however, translating multiple inter‑correlated measurements into diagnostic insight remains challenging. Objective: To present an explainable engineering workflow combining principal component analysis (PCA) and K‑means clustering to identify skeletal Class II phenotypes in adults.

Methods: Sixty‑three CBCT variables from 120 Yemeni adults were standardized, reduced via PCA, and clustered (k = 2–8). Internal validity was evaluated with silhouette and Davies–Bouldin indices and stability across multiple random starts; phenotypes were mapped to clinical strategies.

Results: Seven components explained ≈60% of variance. A five‑cluster solution balanced cohesion, separation, and clinical interpretability, delineating deep‑bite and open‑bite tendencies, mandibular retrusion patterns, and incisor‑protrusive phenotypes. Solutions were stable across seeds and simple demographic strata.

Conclusions: The PCA→K‑means pipeline provides a transparent, reproducible framework for AI‑assisted orthodontic diagnosis and phenotype‑based treatment planning, aligning with biomedical imaging and radiomics workflows.

Dentistry

CBCT

3-D imaging

radiomics

explainable AI

principal component analysis

K‑means clustering

phenotyping

orthodontic diagnostics

Cone‑beam computed tomography (CBCT) has transformed orthodontic imaging by providing 3‑D datasets with isotropic voxels and high geometric fidelity, enabling precise linear and angular assessments across craniofacial regions. Yet, clinically collected CBCT variables are numerous and correlated, complicating etiologic reasoning and treatment planning. From a bioengineering perspective, dimensionality reduction with PCA preserves multivariate structure while K‑means clustering provides an interpretable partitioning of patient phenotypes using Euclidean geometry and centroid summaries. We apply a transparent PCA→K‑means pipeline to adults with skeletal Class II malocclusion and report components, clusters, and phenotype‑specific clinical implications. To align with the Diagnostic Imaging and Radiomics focus, we emphasize explainability, reproducibility, and the potential to integrate into AI‑ready imaging workflows. [7–9] [1–3] [6, 11, 12]

2.1 Study Design and Participants

This retrospective observational study evaluated 120 Yemeni adults diagnosed with skeletal Class II malocclusion. Inclusion criteria comprised complete CBCT records, adult dentition, and absence of syndromic conditions, craniofacial trauma, or prior orthognathic surgery. Exclusion criteria were poor image quality, artifacts interfering with landmarking, and missing key variables.

2.2 Imaging Protocol and Measurements

CBCT scans were acquired on a PaX‑Flex 3D P2 system with a medium field of view. Landmarks and measurements were extracted in Invivo 6.0 following standardized operating procedures. Sixty‑three variables captured sagittal, vertical, and transverse indicators, as well as maxillary and mandibular incisor inclinations and positions. [7–9, 15]

2.3 Preprocessing and Dimensionality Reduction

Variables were screened for completeness, standardized to z‑scores, and subjected to PCA on the correlation matrix. Component retention balanced eigenvalue > 1, visual inspection of the scree profile, cumulative variance criteria, and anatomical interpretability of loadings. [1–3, 10]

2.4 Unsupervised Clustering

K‑means clustering was performed on retained PCA scores using Euclidean distance. Candidate solutions for k ranging from 2 to 8 were fitted with multiple random initializations to avoid local minima. The smallest k exhibiting strong cohesion/separation and clinical interpretability was selected. [6, 11, 12]

2.5 Cluster Validity and Sensitivity Analyses

Silhouette coefficients and Davies–Bouldin indices quantified within‑cluster cohesion and between‑cluster separation. Robustness was assessed across seeds and simple stratifications (e.g., sex). [4, 5]

2.6 Statistical Environment and Reproducibility

Analyses followed reproducible steps implementable in common environments (e.g., Python/R). The pipeline is intentionally simple—standardization, PCA on the correlation matrix, K‑means on retained scores, and validity indices—facilitating multi‑center replication. [1–3]

3.1 Component Loadings and Anatomical Interpretation

Seven principal components captured approximately 60% of variance. PC1 emphasized cranial‑base length; PC2/PC3 captured variance in mandibular plane steepness and facial height; PC4 reflected maxillary sagittal position with secondary contributions from incisor inclination. [7–9, 13–15]

3.2 Cluster Centroids and Clinical Readouts

K‑means on retained scores yielded a five‑cluster solution balancing internal validity and clinical interpretability. Centroid inspection clarified actionable differences: deep‑bite clusters favored torque control and bite‑opening mechanics; open‑bite clusters benefited from vertical control; mandibular retrusion clusters mapped to growth modification or orthognathic pathways depending on age; incisor‑protrusive patterns called for anchorage planning and torque adjustments. [4, 5]

3.3 Sensitivity to Component Retention and k‑Selection

Retaining fewer than seven components modestly reduced silhouette profiles, while adding components beyond seven increased within‑cluster variance without improving separation. k = 4 merged distinct vertical phenotypes, whereas k = 6 split a stable centroid into two near duplicates; thus, k = 5 offered the best balance of parsimony and interpretability. [4, 5]

3.4 Visualization

Two‑dimensional score plots (PC1 vs. PC2) illustrated separable but partially overlapping cloud structures consistent with realistic craniofacial heterogeneity. The workflow and scree/cluster plots are provided as Figs. 1–3.

This work complements deep‑learning efforts in dental imaging by focusing on unsupervised phenotyping and transparent dimensionality reduction, which are essential for explainable decision support. Internal validity indices supported k = 5 and phenotypes were clinically interpretable—a practical criterion often overlooked by purely performance‑driven optimization. Generalizability remains a limitation because the cohort represents Yemeni adults; multi‑center validation across devices/software is warranted. Beyond morphology, integration with outcomes and patient‑reported measures would support phenotype‑to‑protocol pathways and personalized biomechanics. [16–18, 20]

A compact PCA→K‑means pipeline converts high‑dimensional CBCT variables into reproducible skeletal Class II phenotypes suitable for radiomics and clinical workflows. The approach emphasizes dimensionality reduction, unsupervised learning, and objective validation and can be deployed across multi‑center datasets to advance explainable, AI‑assisted orthodontic diagnostics.

5.1 Practical Checklist for Clinical Deployment

• Governance: Verify IRB oversight and de‑identification.

• Protocols: Fix acquisition presets and document scanner settings.

• Measurement: Calibrate operators; perform periodic ICC checks.

• Analytics: Standardize variables; retain components by elbow + interpretability; use multi‑start K‑means; report silhouette/Davies–Bouldin indices.

• Translation: Maintain phenotype templates mapping centroids to orthodontic strategies; monitor outcomes.

• Safety: Apply ALARA/ALADA principles; avoid unnecessary volumetric imaging.

Institutional Review Board Statement

This study was conducted in accordance with the Declaration of Helsinki and approved by the Medical Ethics Committee of Sana’a University, Yemen (Approval ID: OR 19/11/2023).

Informed Consent Statement

Patient consent was waived due to the retrospective design and de‑identified nature of the dataset.

Data Availability Statement

Data are available from the corresponding author upon reasonable request due to ethical restrictions and de‑identification.

Author Contributions

Conceptualization, S.M.B.H.; methodology, S.M.B.H., M.K.J., and M.H.A.-Q.; software/visualization, S.M.B.H.; formal analysis, S.M.B.H.; resources, M.K.J. and M.H.A.-Q.; data curation, S.M.B.H.; writing—original draft, S.M.B.H.; writing—review and editing, S.M.B.H., M.K.J., and M.H.A.-Q.; supervision, S.M.B.H. All authors approved the submitted version.

Funding

This research received no external funding.

Conflicts of Interest

The authors declare no conflict of interest.

Acknowledgments

We thank the Medical Ethics Committee of Sana’a University for oversight and Genesis Medical & Cosmetics Company for administrative support.

Abdi H, Williams LJ. Principal component analysis. WIREs Comput Stat. 2010;2(4):433–59. https://doi.org/10.1002/wics.101
Jolliffe IT, Cadima J. Principal component analysis: a review and recent developments. Philos Trans A. 2016;374(2065):20150202. https://doi.org/10.1098/rsta.2015.0202
Hastie T, Tibshirani R, Friedman J. The Elements of Statistical Learning. 2nd ed. Springer; 2009. https://link.springer.com/book/10.1007/978-0-387-84858-7
Rousseeuw PJ. Silhouettes: A graphical aid to the interpretation and validation of cluster analysis. J Comput Appl Math. 1987;20:53–65. https://doi.org/10.1016/0377-0427(87)90125-7
Davies DL, Bouldin DW. A cluster separation measure. IEEE TPAMI. 1979;1:224–27. https://doi.org/10.1109/TPAMI.1979.4766909
Hartigan JA, Wong MA. A K‑Means clustering algorithm. J R Stat Soc C. 1979;28(1):100–8. https://doi.org/10.2307/2346830
Scarfe WC, Farman AG. What is cone‑beam CT and how does it work? Dent Clin North Am. 2008;52(4):707–30. https://doi.org/10.1016/j.cden.2008.05.005
Venkatesh E, Elluru SV. Cone beam computed tomography: basics and applications in dentistry. J Istanb Univ Fac Dent. 2017;51(Suppl 1):S102–21. https://pmc.ncbi.nlm.nih.gov/articles/PMC5750833/
Pauwels R, Beinsberger J, Collaert B, et al. Effective dose range for dental CBCT scanners. Eur J Radiol. 2012;81(2):267–71. https://doi.org/10.1016/j.ejrad.2010.11.028
Izenman AJ. Modern Multivariate Statistical Techniques. Springer; 2008. https://link.springer.com/book/10.1007/978-0-387-78189-1
MacQueen J. Some methods for classification and analysis of multivariate observations. In: Proc 5th Berkeley Symp Math Stat Prob. 1967:281–97. https://projecteuclid.org/journals/proceedings-of-the-fifth-berkeley-symposium-on-mathematical-statistics-and-probability/volume-1/issue-1/Some-methods-for-classification-and-analysis-of-multivariate-observations/bsmsp/1200512992.full
Lloyd SP. Least squares quantization in PCM. IEEE Trans Inf Theory. 1982;28(2):129–37. https://doi.org/10.1109/TIT.1982.1056489
Damstra J, Fourie Z, Huddleston Slater JJ, Ren Y. Accuracy of linear measurements from CBCT‑derived 3‑D surface models. Am J Orthod Dentofacial Orthop. 2010;137(4):S50–S1. https://pubmed.ncbi.nlm.nih.gov/20122425/
Gribel BF, Gribel MN, Manzi FR, Brooks SL, McNamara JA Jr. Accuracy and reliability of craniometric measurements on lateral cephalograms vs CBCT. Angle Orthod. 2011;81(1):26–35. https://media.dent.umich.edu/labs/mcnamara/files/Accuracy%20and%20reliability%20of%20craniometric%20measurements%20on%20lateral%20cephalograms%20and%20CBCT.pdf
Scarfe WC, Aboelsaad N, Farman AG, et al. CBCT in orthodontics. Aust Dent J. 2017;62(S1):33–50. https://onlinelibrary.wiley.com/doi/abs/10.1111/adj.12479
Schwendicke F, Samek W, Krois J. Artificial intelligence in dentistry: chances and challenges. J Dent Res. 2020;99(7):769–74. https://doi.org/10.1177/0022034520915714
Arsiwala‑Scheppach LT, Kronig J, et al. Machine learning in dentistry: a scoping review. J Clin Med. 2023;12(3):937. https://pubmed.ncbi.nlm.nih.gov/36769585/
Schwendicke F, Chaurasia A, Arsiwala‑Scheppach L, et al. Deep learning for cephalometric landmark detection: systematic review and meta‑analysis. Clin Oral Investig. 2021;25:4299–4312. https://link.springer.com/article/10.1007/s00784-021-03990-w
Pauwels R, Smeets B, et al. Converting DAP to effective dose for dental CBCT. Physica Medica. 2023;110:102–10. https://www.physicamedica.com/article/S1120-1797(23)00116-3/fulltext
Lakhotia S, et al. Machine learning in dentistry: scoping review. PLOS Digit Health. 2025;4(9):e0000940. https://doi.org/10.1371/journal.pdig.0000940

The authors declare no competing interests.

GraphicalAbstract.png
Graphical Abstract. AI‑CBCT phenotyping workflow.

Download PDF

Version 1

posted

You are reading this latest preprint version

CBCT-Based Morphometric Clustering for Orthodontic Diagnostics Using PCA and K-Means: An Explainable AI Workflow

Status:

Version 1

Abstract

Figures

1. Introduction

2. Materials and Methods

2.1 Study Design and Participants

2.2 Imaging Protocol and Measurements

2.3 Preprocessing and Dimensionality Reduction

2.4 Unsupervised Clustering

2.5 Cluster Validity and Sensitivity Analyses

2.6 Statistical Environment and Reproducibility

3. Results

3.1 Component Loadings and Anatomical Interpretation

3.2 Cluster Centroids and Clinical Readouts

3.3 Sensitivity to Component Retention and k‑Selection

3.4 Visualization

4. Discussion

5. Conclusions

5.1 Practical Checklist for Clinical Deployment

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1