Related Experiment Video
Updated: Aug 22, 2026
Automated Joint Space Detection Improves Bone Segmentation Accuracy
Published on: November 28, 2025
Two-stage deep learning framework for magnetic resonance imaging-based detection of temporomandibular joint
Glenda Tsai1, Dongwon Park1, Kaan Orhan2
1Schulich School of Medicine and Dentistry, Western University, London, ON, Canada.
Objectives:
Cone-beam computed tomography (CBCT) is the reference standard for detecting osseous changes in temporomandibular joint osteoarthritis (TMJ-OA) but involves ionizing radiation. MRI avoids radiation exposure and enables simultaneous soft and hard tissue evaluation; however, its sensitivity for osseous abnormalities remains limited. This study aimed to develop and evaluate a proof-of-concept two-stage deep learning framework for automated MRI-based TMJ-OA detection.
Methods:
A retrospective dataset of 200 bilateral TMJ MRI examinations (400 condyles; 113 osteoarthritic, 287 non-osteoarthritic) was collected from two academic dental institutions. Stage 1 employed a U-Net architecture for automated mandibular condyle segmentation using manually annotated masks. Stage 2 applied ResNet-50 and ResNet-101 to classify segmented condylar regions as absent, normal, or osteoarthritic. Labels were established by two experienced oral and maxillofacial radiologists using CBCT-confirmed osseous findings and DC/TMD clinical criteria. Patient-level partitioning (70/15/15%) and class-weighted loss functions addressed data leakage and class imbalance. An experienced radiologist and a newly graduated clinician provided human benchmark comparisons on the held-out test subset.
Results:
U-Net segmentation achieved a Dice similarity coefficient of 0.944, Intersection over Union of 0.894, and AUC of 0.97, with balanced precision and recall of 0.948. ResNet-50 yielded an F1 score of 0.813 (precision 0.867, recall 0.765); ResNet-101 yielded 0.800 (precision 0.778, recall 0.824). The experienced radiologist achieved an F1 score of 0.970; the newly graduated clinician achieved 0.414.
Conclusions:
This study demonstrates the feasibility of a segmentation-first, classification-second deep learning pipeline for automated MRI-based TMJ-OA detection, providing a methodological foundation for future multi-institutional validation of a radiation-free diagnostic approach to TMJ-OA.
