Evaluation of supervised machine learning models in predicting temporomandibular joint disc displacement on 3T magnetic resonance imaging

OBJECTIVE: To evaluate supervised machine learning (ML) models for classifying temporomandibular joint (TMJ) disc displacement on MRI using morphometric and signal intensity features. METHODS: This retrospective study analyzed 324 TMJs from 162 individuals who underwent 3T MRI. Extracted features included condylar anteroposterior and mediolateral diameters, disc and condyle morphology, and lateral pterygoid muscle signal intensity ratios. Six ML algorithms (Random Forest, Gaussian Naïve Bayes, Logistic Regression, AdaBoost, Gradient Boost, and k-Nearest Neighbor) were evaluated using stratified 5-fold cross-validation. Performance was assessed with accuracy, macro-averaged recall, precision, F1-score, and ROC-AUC. RESULTS: All models demonstrated good classification performance (ROC-AUC >0.80). AdaBoost achieved the highest ROC-AUC (0.88), while Gaussian Naïve Bayes showed the most balanced overall metrics. Mediolateral condylar diameter and disc morphology were key features associated with disc displacement categories. CONCLUSIONS: ML models can identify MRI-based morphometric patterns related to TMJ disc displacement and may support radiologic assessment, while clinical diagnosis should continue to rely on established standards of care.