Diagnostic Performance of Commercial Large Vision Models in Assessing Temporomandibular Joint Disorders Using Isolated Panoramic Temporomandibular Joint View Radiographs: A Cross-Sectional Study
Mandavi S. Waghmare, Ronak A. Varyani, Sandeep S. Pagare, Utkarsha U. Sawant, Nisha N. Bharos, Chintan ShahAbstract
Background:
Artificial intelligence (AI) and large vision models (LVMs) are increasingly explored as diagnostic aids in oral radiology. Diagnosing temporomandibular disorders on two-dimensional temporomandibular joint (TMJ) open–close radiographs remains challenging because of subtle osseous changes and anatomical superimposition. Comparative evidence evaluating contemporary commercial LVMs against expert interpretation is limited.
Objective:
To compare the diagnostic performance of five commercial LVMs with an expert reference standard for interpreting 2D TMJ open–close radiographs.
Methods:
A retrospective study analyzed 71 anonymized TMJ open–close radiographs. All images were assessed using a validated 12-point diagnostic proforma, which also served as standardized prompts for the LVMs. An oral radiology expert with over 20 years of experience established the reference standard. Diagnostic agreement was evaluated using Cohen’s kappa, accuracy, and confusion matrices.
Results:
OpenAI achieved the highest agreement (
Conclusion:
Among the evaluated LVMs, the OpenAI vision-enabled model showed the strongest agreement with expert TMJ radiograph interpretation.