Advantages of deep learning with convolutional neural network in detecting disc displacement of the temporomandibular joint in magnetic resonance imaging

Jonghyun Won, Yeon-Hee Lee, Seunghyeon Kim, Q.-Shick Auh, Yung-Kyun Noh

Abstract:

This study investigated the usefulness of deep learning-based automatic detection of anterior disc
displacement (ADD) from magnetic resonance imaging (MRI) of patients with temporomandibular
joint disorder (TMD). Sagittal MRI images of 2520 TMJs were collected from 861 men and 399 women
(average age 37.33 ± 18.83 years). A deep learning algorithm with a convolutional neural network was
developed. Data augmentation and the Adam optimizer were applied to reduce the risk of overfitting
the deep-learning model. The prediction performances were compared between the models and
human experts based on areas under the curve (AUCs). The fine-tuning model showed excellent
prediction performance (AUC = 0.8775) and acceptable accuracy (approximately 77%). Comparing the
AUC values of the from-scratch (0.8269) and freeze models (0.5858) showed lower performances of the
other models compared to the fine-tuning model. In Grad-CAM visualizations, the fine-tuning scheme
focused more on the TMJ disc when judging ADD, and the sparsity was higher than that of the fromscratch
scheme (84.69% vs. 55.61%, p < 0.05). The three fine-tuned ensemble models using different
data augmentation techniques showed a prediction accuracy of 83%. Moreover, the AUC values of
ADD were higher when patients with TMD were divided by age (0.8549–0.9275) and sex (male: 0.8483,
female: 0.9276). While the accuracy of the ensemble model was higher than that of human experts,
the difference was not significant (p = 0.1987–0.0671). Learning from pre-trained weights allowed the
fine-tuning model to outperform the from-scratch model. Another benefit of the fine-tuning model
for diagnosing ADD of TMJ in Grad-CAM analysis was the deactivation of unwanted gradient values
to provide clearer visualizations compared to the from-scratch model. The Grad-CAM visualizations
also agreed with the model learned through important features in the joint disc area. The accuracy
was further improved by an ensemble of three fine-tuning models using diversified data. The main
benefits of this model were the higher specificity compared to human experts, which may be useful for
preventing true negative cases, and the maintenance of its prediction accuracy across sexes and ages,
suggesting a generalized prediction.