Two-stage deep learning approach for screening disk displacement of the temporomandibular joint using orthopantomogram

This study aimed to develop a deep learning-based approach for the initial screening of disk displacement (DD) of the temporomandibular joint (TMJ) using orthopantomograms (OTG). A two-stage deep learning model was proposed: first, regions of interest were detected from 1,469 orthopantomogram images using YOLOv5s; second, DD status was identified from 1,564 joints with the magnetic resonance imaging-verified DD status using ResNet-18 and DANet. Diagnostic performance was evaluated through multi-class and binary classification analyses. In three-class classification (normal disk position, DD with reduction, and DD without reduction), the model achieved an overall accuracy of 71.86%. It performed well in identifying normal disk position (F1 score: 78.57%) and DD without reduction (F1 score: 80.99%) but showed lower performance in detecting DD with reduction (F1 score: 43.84%). In the binary classification, where DD with reduction and DD without reduction were combined into a single class, the model demonstrated improved accuracy (82.04%), sensitivity (78.85%), and specificity (87.30%). The model exhibited potential in estimating DD severity (R² = 0.4871). Given its strong ability to differentiate between normal disk position and DD in a binary classification setting, this tool has the potential to serve as an initial screening tool for TMJ DD in clinical practice.