Automated Fractal Analysis of Right and Left Condyles on Digital Panoramic Images Among Patients With Temporomandibular Disorder (TMD) and Use of Machine Learning Algorithms in the Diagnosis of TMD

TL;DR

This study uses fractal analysis and machine learning to distinguish TMD patients from healthy individuals using panoramic dental images.

Contribution

The study introduces automated fractal analysis and XGBoost machine learning for TMD diagnosis using digital panoramic images.

Findings

TMD patients had significantly lower fractal dimension values compared to healthy controls.

XGBoost achieved an ROC-AUC of 0.9449, showing strong diagnostic performance for TMD.

Gender and age influenced fractal dimensions, with the most notable differences in the 30-39 age group.

Abstract

Background & purpose

The diagnosis of temporomandibular disorder (TMD) can be a challenging and arduous task. The role of advanced imaging modalities, such as computed tomography & magnetic resonance imaging, in the evaluation of temporomandibular joints is indispensable. However, panoramic imaging remains a widely available and cost-effective imaging modality in many clinical setups. This study aims to automate fractal analysis of mandibular condyles on panoramic radiographs using ImageJ macros code and to compare fractal dimensions between TMD patients and age-group-matched healthy controls. Additionally, the study evaluates age- and gender-related variations in fractal dimensions and investigates the utility of machine learning classifiers for differentiating patients with TMD from healthy individuals.

Materials and methods

A total of 220 subjects were selected for the study, with 110 patients with TMD & 110 healthy controls. Additionally, they were divided into four age groups: 18-29 years, 30-39 years, 40-49 years, and ≥ 50 years. Right & left condyles on 220 digital orthopantomogram (OPG), of which 110 belonged to patients with TMD and 110 to age group-matched healthy controls, were analyzed with ImageJ 1.42q for the automated calculations of fractal dimensions (FD) with the box counting method. Data obtained was utilized to train (70%), validate (15%), and test (15%) six machine learning (ML) algorithms: random forest classifier, logistic regression, support vector machine (SVM), gradient boosting, K-Nearest Neighbors & XGBoost to differentiate between patients with TMD & healthy controls based on the features such as FD values and age & gender variables.

Results

For right & left condyles, the FD distribution showed a similar pattern across most of the age groups, except for the control group between the ages of 18 and 29 years. FD values for the patients with TMD were statistically lower in most of the age groups compared to those of the controls, with the greatest visual differences observed in the 30-39 age group and the female participants. Overall mean FD value for patients was 1.2232 and controls was 1.2944. For ML, the best performing model was XGBoost with the validation F1 Score: 0.9620, test F1 Score: 0.8750, test accuracy: 0.8936, test precision: 0.8750, test recall: 0.8750, and test receiver operating characteristic area under the curve (ROC AUC): 0.9449 for discriminating between patients with TMD & healthy controls.

Conclusion

Patients with TMD showed highly statistically lowered FD values for the mandibular condyles compared to those of the age group-matched controls, with significant overall and group-wise gender differences. FD values show robust prognostic power for differentiating patients with TMD from the respective controls. The application of a ML algorithm, i.e., the XGBoost, attained an ROC-AUC of 0.9449, indicating excellent diagnostic performance, which validates the capability of FD as a useful diagnostic radiographic marker.

Figures