Skip to main navigation Skip to search Skip to main content

Computer-Aided Analysis of 3D Dental Models for Cleft Lip and Palate Assessment

  • Yihang Xie

Student thesis: Doctoral Thesis › Doctor of Philosophy

Abstract

Digital dentistry has increasingly adopted three-dimensional (3D) imaging and computational methods for diagnosis, treatment planning, and outcome assessment in craniofacial care. In the management of cleft lip and palate (CLP), 3D digital dental models offer a more objective and comprehensive basis for clinical evaluation compared to conventional approaches. The principal surrogate measure of facial growth in children born with CLP is maxillary arch constriction, which in routine practice is judged by visual evaluation of the form of the maxillary arch and of the occlusal relationship between the maxillary and mandibular teeth. However, established clinical scoring and classification tools for evaluating dental arch relationships and documenting alveolar defects still rely heavily on manual interpretation and lack automated support leveraging 3D digital models, making them time-consuming and susceptible to subjective bias. Moreover, existing automated methods for 3D dental model analysis, including model orientation, tooth segmentation, and landmark detection, have been developed primarily for regular, non-cleft dental anatomies and often perform poorly on cleft-affected dentitions due to their complex morphological abnormalities.

This thesis addresses these gaps by developing a series of computational methods for the automated analysis of 3D dental models in CLP assessment. Four closely related problems are investigated: (1) robust model orientation, (2) tooth segmentation under complex dentition conditions, (3) automated landmark detection for dental arch assessment, and (4) automated classification of neonatal alveolar bone defects. For orientation, a geometry-aware method is proposed that combines principal component analysis (PCA) with a minimum-area bounding box to determine axis directions. For tooth segmentation, three approaches are developed, ranging from adaptive region growing and harmonic-field-based segmentation to a deep learning framework incorporating contrastive boundary learning and curvature-enhanced boundary delineation. For landmark detection, tooth-type-specific strategies are designed, including watershed-based cusp detection for canines and molars and curvature-guided incisal edge recovery for incisors. For alveolar defect classification, a method is proposed that segments the alveolar bone into anatomical regions and characterizes each defect by estimating its geometry from a surface curve spanning the cleft gap, thereby enabling classification of laterality and completeness.

The proposed orientation method achieves valid orientation rates above 90\% across all three datasets, substantially outperforming direct PCA, which drops to as low as 5\% on neonatal cleft models. For tooth segmentation, the proposed geometric methods improve segmentation accuracy and tooth detection over fixed-parameter baselines but remain limited on severely deformed cleft models. The deep learning-based framework proves robust to the complex morphological variations in cleft-affected models and outperforms both the geometric methods proposed in this thesis and several representative deep learning approaches across all evaluation metrics. For dental landmark detection, the proposed method achieves correct identification rates exceeding 95\% across all tooth types, including incisal edge midpoints on incisors, cusp tips on canines, and buccal cusps on molars. The alveolar defect classification method correctly classifies 18 out of 20 neonatal maxillary models covering a range of cleft types and severities, demonstrating the feasibility of automated alveolar cleft classification from 3D surface data.

This thesis contributes automated methods for 3D dental model analysis specifically tailored to the challenges of cleft lip and palate. The proposed methods address several key tasks, including model orientation, tooth segmentation, landmark detection, and defect classification. The modular design of these methods enables their integration into an end-to-end pipeline, providing a technical foundation for standardized, reproducible assessment of 3D oral data, while reducing reliance on manual clinical interpretation.
Date of Award2026
Original languageEnglish
Awarding Institution
  • University of Dundee
SponsorsChina Scholarship Council
SupervisorPing Lin (Supervisor)

Keywords

  • 3D dental model analysis
  • Tooth Segmentation
  • Deep Learning
  • Cleft lip and palate
  • Pose normalization

Cite this

'