Automatic Prediction of TMJ Disc Displacement in CBCT Images Using Machine Learning

Area of Science:

• Dentistry
• Radiology
• Artificial Intelligence

Background:

• Temporomandibular joint (TMJ) disc displacement diagnosis relies heavily on MRI, which presents accessibility challenges due to cost and practicality.
• Cone-beam computed tomography (CBCT) is more accessible but lacks the diagnostic detail of MRI for TMJ disc displacement.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting TMJ disc displacement using CBCT-based radiomics features.
• To assess the feasibility of replacing MRI with CBCT and ML for diagnosing TMJ disc displacement, particularly disc displacement without reduction (DDWOR).

Main Methods:

• Radiomics features were extracted from CBCT images of 247 mandibular condyles from 134 patients who also had MRI scans.
• Two ML models, Random Forest (RF) and XGBoost, were trained and compared across three classification experiments.
• Experiments involved classifying normal, disc displacement with reduction (DDWR), and disc displacement without reduction (DDWOR) states.

Main Results:

• The Random Forest (RF) model outperformed XGBoost in all experiments.
• Experiment 3, differentiating DDWOR from other conditions, achieved the highest AUC of 0.86 (RF) and 0.85 (XGBoost).
• Classifying all three groups (Experiment 1) yielded the lowest accuracy, with AUCs of 0.63 (RF) and 0.59 (XGBoost).

Conclusions:

• ML models utilizing CBCT radiomics show promise for predicting TMJ disc displacement.
• This AI-driven approach can serve as an assistive tool, particularly for identifying DDWOR, which requires careful management.
• CBCT-based radiomics offers a potential, more accessible alternative to MRI for diagnosing TMJ disc displacement.