Source-linked AI summary
Morphological Decoupling-Based Skeletal Classification for Clinical Assessment of Malocclusion
Zhichun Jin, Zhicheng He, Hao Xu, Dongyang Li, Lin Wang, Hongliang Ren, Long Bai
TL;DR
Manual CBCT-based malocclusion grading is time-consuming and subject to inter-operator variability. TeethGNN predicts morphological indicators from CBCT images, fuses them with image features using a GNN, and applies collaborative calibration; experiments report strong performance, while the single-center dataset and omitted functional parameters limit scope.
Problem
Manual CBCT-based skeletal grading requires time-consuming expert measurement and is subject to inter-operator variability, while GNN outputs can be poorly calibrated.
Method
TeethGNN predicts morphological indicators from a single CBCT-reconstructed lateral cephalogram, fuses them with image features using a GNN, and applies explicit adversarial and implicit topological calibration.
Results
TeethGNN achieves strong overall performance against state-of-the-art methods, with the framework and its calibration components supported by extensive experiments and ablations.
Takeaways & Limitations
The system provides a vision-based CBCT measurement approach with potential to support faster, more consistent computer-aided orthodontic diagnosis.
Takeaways & Limitations
The study uses a single-center dataset and does not incorporate dynamic functional parameters such as bite force or soft-tissue dynamics.
Abstract
from arXiv · showhide
Malocclusion skeletal grading is a fundamental task in orthodontics, critical for diagnosis and treatment planning. Traditionally, cone-beam computed tomography (CBCT) is used for visual measurement, and the reconstructed lateral cephalograms are handed over to expert dentists for diagnosis. However, manual review is time-consuming, labor-intensive, and subject to inter-operator variability. Therefore, an automatic CBCT-based system is needed for reliable malocclusion skeletal grading. In this case, we develop TeethGNN, a novel graph-based framework designed to combine CBCT image features with morphological information for accurate and efficient malocclusion grading. TeethGNN utilizes a decoupled learnable decoder to directly predict key morphological indicators from CBCT images, eliminating the need for manual measurements. These morphological features are then fused with image features using a graph neural network (GNN), which effectively models the relationships between the modalities. To further enhance robustness and calibration, we introduce a collaborative calibration strategy. This strategy combines multi-scale graph adversarial perturbation for explicit calibration and nonlinear topological graph calibration for implicit confidence adjustment. Extensive experiments and ablation studies on our collected clinical dataset demonstrate that our malocclusion measurement system achieves 77.08\% in accuracy and 89.61\% in AUC, outperforming the compared state-of-the-art methods. These results validate the effectiveness of graph-based multimodal fusion and collaborative calibration in improving malocclusion grading performance. Our system shows strong potential for advancing computer-aided orthodontic diagnosis, providing an accurate and reliable solution for vision-based clinical measurement and diagnosis.
1. Introduction
TeethGNN addresses time-consuming and variable manual CBCT-based skeletal grading by decoupling morphological prediction from image features and fusing both through a calibrated graph model.
- 1. Introduction: Manual CBCT grading requires expert visual inspection and measurements of reconstructed lateral cephalograms, creating time demands and inter-operator variability.Skeletal grading categorizes cases into Classes I, II, and III according to maxilla–mandibula position.
- 1. Introduction: GNN message passing models relationships between morphological and image nodes, preserving numerical morphological information during multimodal fusion.The approach uses separate graph nodes for decoupled morphological information and image representations.
- 1. Introduction: GNN confidence calibration is motivated by mismatches between predicted probabilities and actual correctness, which can undermine threshold-based clinical decisions.The paper links miscalibration to graph structure, neighboring information, shallow architectures, and smaller training datasets.
- 1. Introduction: The collaborative calibration strategy injects multi-scale adversarial perturbations into graph-node embeddings and applies nonlinear topological calibration to prediction confidence.The perturbation operates without changing graph topology, while the topological stage makes neighboring-node confidence more consistent.
- 1. Introduction: The system combines CBCT image representations with predicted morphological information through a graph-based framework for malocclusion skeletal grading.A decoupling learnable decoder predicts ANB and MP-FH indicators from the reconstructed lateral cephalogram, avoiding manually supplied morphological measurements.
- 1. Introduction: Extensive comparative and ablation experiments report superior malocclusion skeletal-grading performance over existing classification solutions and improvements from the calibration design.The supplied passages characterize the system as a promising CBCT-based computer-aided diagnostic measurement approach.
2. Related Work
Related work applies AI, CNNs, and graph learning to orthodontic measurement, while highlighting limitations involving data scale, landmark direction, dimensionality, and input variability.
- 2. Related Work: Other AI-based oral measurement work has used bite-force information with random forests, illustrating the range of data modalities explored for malocclusion-related assessment.The cited bite-force system achieved 87.83% accuracy.
- 2. Related Work: CNN-based malocclusion methods have achieved high landmark-detection rates and low point-to-point errors on cephalograms.The cited examples report landmark detection rates exceeding 98% and point-to-point error as low as 1.26mm.
- 2. Related Work: Existing CNN-based grading methods remain limited by small samples, inability to predict landmark-change direction, reliance on 2D images, and variable landmark definitions and input quality.Reliance on lateral cephalograms omits 3D information.
- 2. Related Work: Graph representation learning has been applied in biomedicine to model gene interactions, drug-related relationships, adverse reactions, and personalized treatment recommendations.These applications use heterogeneous or large-scale biomedical data to support prediction and recommendation tasks.
- 2. Related Work: Dental GNN applications include non-Euclidean tooth segmentation and adaptive graph modeling of tooth arrangement in orthodontic treatment.Reported targets include robustness to malocclusions and scanning noise, as well as complex extraction and overcrowding cases.
3. Methods
TeethGNN processes CBCT-derived images through complementary visual and morphological pathways, fuses their representations with a graph network, and applies collaborative calibration for robust diagnosis.
- 3.1. AI-assisted System: The system uses CBCT scans to reconstruct lateral cephalograms that are processed into diagnostic references for dentists.Full-cranial imaging is configured with a 15 × 15 cm field of view and produces lateral cephalograms for TeethGNN inference.
- 3.3. Multi-modality Teeth Graph: The image pathway uses ResNet50 to extract global visual features, while separate encoder-decoder pairs predict ANB and MP-FH morphological angles.The morphological pathway is trained with mean squared error so predicted angles match their ground-truth values.
- 3.3.1. Multi-modality Teeth Graph: During warm-up, the image encoder learns morphological information with MSE loss before full training jointly optimizes graph-based multimodal representation learning.The framework distinguishes trainable and frozen modules across the warm-up and full-training stages.
- 3.3.1. Multi-modality Teeth Graph: TeethGNN combines predicted morphological measurements and image features in a graph representation for malocclusion diagnosis.The graph contains image, ANB, and MP-FH nodes whose embeddings are iteratively updated to capture their interactions.
- 3.3.2–3.3.3. Collaborative Calibration: Multi-scale graph adversarial perturbations modify node embeddings without changing graph topology, improving robustness and explicitly calibrating predictions.Perturbations are optimized by projected gradient ascent and combined with nonlinear topological calibration that regularizes confidence across connected nodes.
- 3.3.4. Training and Inference Recipe: The full objective combines classification loss with calibration regularization so the model uses image and predicted morphological information without ground-truth morphology at inference.The hyperparameter α balances classification accuracy against calibration smoothness.
4. Results
TeethGNN outperformed CNN, Transformer, foundation-model, and graph-based baselines, while calibration and fusion ablations support its design choices.
- 4.3. Experimental Results: 0.7708 accuracy and 0.6516 F1-score exceeded ResNet50 by 0.0208 and 0.0916, respectively, despite ResNet50 being the strongest CNN baseline.
- 4.3. Experimental Results: 0.7708 accuracy and 0.8961 AUC made TeethGNN the strongest overall model, exceeding the compared graph, CNN, Transformer, and foundation-model baselines.TeethGNN also achieved a 0.6516 F1-score, with the strongest graph-based comparison showing 0.6667 accuracy and 0.4500 F1-score.
- 4.3. Experimental Results: 0.8961 AUC was the highest reported value, while TeethGNN exceeded VisionGNN by 0.1041 accuracy and 0.2016 F1-score.
- 4.4. Ablation Studies: 0.0910 ECE and 0.4987 Brier score were achieved by the complete MGAP-plus-NTGC model, which also retained the best accuracy, F1-score, and AUC.The results suggest that both calibration components improve confidence alignment without compromising classification performance.
- 4.4. Ablation Studies: 0.8125 accuracy was obtained at α = 0.05, whereas α = 1 produced the highest precision of 0.8178 and F1-score of 0.7171 but lower accuracy and AUC.The ablation identifies a trade-off between the primary classification objective and calibration loss.
- 4.4. Ablation Studies: 0.7708 accuracy and 0.8961 AUC were the best fusion results, with graph-based fusion exceeding self-attention and other multimodal strategies across the reported metrics.The passage attributes this pattern to modeling relations between modalities and capturing complementary information.
5. Discussion
TeethGNN shows promise as a CBCT-based decision-support system for malocclusion grading, while its clinical interpretation remains bounded by error costs, ambiguous cases, missing functional data, radiation considerations, and single-center validation.
- Effectiveness and Clinical Relevance: TeethGNN supports clinical decision-making by combining graph-based image and morphological features, with potential for screening, referral, consistent planning, and longitudinal monitoring.The system is positioned as a decision-support tool rather than an autonomous diagnostic system.
- Limitations and Radiation Considerations: Calibration improves confidence reliability so low-confidence or borderline predictions can be flagged for specialist review instead of acted upon directly.This is particularly relevant because cross-type confusions, especially involving Class III, can correspond to different treatment strategies.
- Limitations and Radiation Considerations: Clinically ambiguous cases remain challenging, including threshold-adjacent ANB values, subtle Class II division distinctions, and underrepresented classes that also generate human-expert disagreement.These cases limit the extent to which the system should replace specialist judgment.
- Limitations and Radiation Considerations: The study does not incorporate dynamic functional parameters such as bite force or soft-tissue dynamics, which are relevant to comprehensive diagnosis.Integrating these modalities is described as challenging because orthodontic conditions are physiologically complex.
- Limitations and Radiation Considerations: The single-center dataset limits diversity in devices, acquisition protocols, and patient populations, so the reported evidence supports internal feasibility rather than broad generalizability.External validation on independent, multicenter cohorts is identified as an essential next step.
- Limitations and Radiation Considerations: CBCT remains necessary for the current skeletal task because ANB and MP-FH depend on bony landmarks unavailable from facial surface scans, so TeethGNN is intended for clinically indicated scans rather than additional radiation.Radiation-free 3D imaging remains promising for soft-tissue evaluation and follow-up.
6. Conclusion
The paper presents TeethGNN as a graph-based system for malocclusion skeletal grading, combining morphological information with CBCT image features. Its experiments and ablations show strong performance and support the effectiveness of its calibration components.
- TeethGNN integrates morphological information and CBCT image features through a graph-based framework for malocclusion skeletal grading.
- The framework uses multi-scale graph adversarial perturbation and nonlinear topological graph calibration to improve robustness, calibration, and diagnostic accuracy.
- Experiments and ablation studies demonstrate strong overall performance against state-of-the-art methods and support the effectiveness of the system’s individual components.
- Future work will expand the dataset, develop advanced graph-based multimodal integration, and adapt the system to broader medical-imaging and automatic-diagnosis applications.
Author Contribution Statement
The author contribution statement assigns responsibilities across conceptualization, data curation, analysis, methodology, software, visualization, writing, and project supervision.
- Zhichun Jin led conceptualization, data curation, formal analysis, investigation, methodology, visualization, and original-draft writing.
- Zhicheng He contributed to conceptualization, data curation, formal analysis, methodology, software, validation, and original-draft writing.
- Hao Xu contributed data curation, visualization, software, and review and editing, while Dongyang Li contributed data curation and visualization.
- Lin Wang contributed conceptualization, funding acquisition, project administration, resources, supervision, and writing and editing.