Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Presentation Icons

Ticketed <br><i>This session/event requires a ticket.</i>

Ticketed
This session/event requires a ticket.

In-person only<br><i>This session is offered in-person only.</i>

In-person only
This session is offered in-person only.

On-demand<br><i>On-demand session available beginning Sept. 15.</i>

On-demand
On-demand session available beginning Sept. 15.

Recorded<br><i>This session will be recorded and available beginning Oct. 6.</i>

Recorded
This session will be recorded and available beginning Oct. 6.

Sold Out<br><i>This ticketed session/event is sold out.</i>

Sold Out
This ticketed session/event is sold out.

63 Views

TMJ

OMS and Other Dental Specialists

Session: Oral Abstract Session Two

SA2 - Utilization of a Residual Neural Network Model for Identification of Temporomandibular Joint Disk Position and Displacement

Friday, September 19, 2025

2:25 PM - 2:35 PM EDT

Location: Room 204C, Walter E. Washington Convention Center

Oral Abstract Presenter(s)

Sayyed Ourmazd Mohseni, DDS

Resident Physician
Parkland Memorial Hospital, UT Southwestern Medical Center
Dallas, Texas

Co-Author(s)

Thomas Schlieve, DDS, MD, FACS

Associate Professor and Graduate Program Director
Parkland Memorial Hospital, UT Southwestern Medical Center
Dallas, Texas

Abstract:
Statement of Problem: Temporomandibular joint (TMJ) disorders have a reported estimated prevalence of 5-12% in the general adult population. The advent of modern imaging modalities has provided an accurate and easily accessible form of non-invasive diagnosis. Artificial intelligence has been a recent avenue of investigation in medicine with promising outcomes for diagnosis and treatment planning. Various types of neural networks have been previously utilized to identify pathologies in TMJ imaging with varying levels of accuracy. The aim of this study was to utilize a modified residual neural network (ResNet) for automated identification of the disk and displacement in pathological conditions.
Materials and

Methods: A total of 50 de-identified temporomandibular joint magnetic resonance imaging (MRI) with contrast were selected for the study. Inclusion criteria included patients presenting with temporomandibular joint pain without resolution of pain from conservative treatment or Botox administration. Radiologist annotations were utilized via transfer learning for training of a modified ResNet-152 model on PyTorch. 60% (n = 30) datasets were utilized for model training followed by 40% (n = 20) for validation and performance evaluation. For evaluation of model performance, the radiologist annotations of disk location and displacement were noted as ground truth and utilized as the current gold standard. After the training period, radiologist annotations were compared to automated annotations from the ResNet model.

Results: After completing the training dataset, the performance dataset was independently annotated by the neural network. Results were then compared to the radiologist reads (gold standard) to calculate Sensitivity (recall), accuracy and F-1 score for detection of disk displacement. Disk structure identification accuracy was assessed via calculation of Intersection over Union (IoU) and Dice coefficients. Cohen’s Kappa (κ) was calculated for agreement between the gold standard and model output. Upon conclusion of data analysis, sensitivity was calculated as 0.95, accuracy at 0.92 with an F-1 score of 0.92. For disk position, an IoU of 0.88 with a dice coefficient of 0.94.

Conclusions: The study demonstrated that a modified residual neural network (ResNet-152) can be trained to identify the disk location with possible anterior displacement accurately when compared to the current gold standard. Current limitations of the study include a small sample size and limited computational capacity. While fundamental at this stage, the study provides insight into increased utilization of artificial intelligence modalities in diagnosis and presurgical planning for treatment of temporomandibular joint disorders. Continued development of such modalities can lead to more streamlined surgical planning with a potential improvement in accuracy.

References

Emhoffs R, Brandlmaier I, Bertram S, et al. Comparing methods for diagnosing temporomandibular joint disc displacement without reduction. J Am Dent Assoc 2002;133(4):442-451.

Jha N, Lee KS, Kim YJ. Diagnosis of temporomandibular disorders using artificial intelligence technologies: A systematic review and meta-analysis. PLoS One. 2022 Aug 18;17(8):e0272715. doi: 10.1371/journal.pone.0272715. PMID: 35980894; PMCID: PMC9387829.