Advantages of Deep Learning with Convolutional Neural Network in Detecting Disc Displacement of the Temporomandibular Joint in Magnetic Resonance Imaging

Abstract

Aim: This study aimed to investigate the usefulness of deep-learning-based automatic detection of anterior disc displacement (ADD) in patients with temporomandibular joint disorder (TMD) using magnetic resonance imaging (MRI). Methods: Sagittal MRI images of 2520 TMJs were collected from the study population (861 men, 399 women; average age 37.33±18.83 years). A deep learning algorithm with a convolutional neural network (CNN) was performed. Data augmentation and the Adam optimizer were applied to reduce the risk of overfitting the deep learning model. The prediction performance of the models and human experts was compared using area under the curve (AUC). Results: The prediction performance of the fine-tuning model was excellent with an AUC of 0.8775, with acceptable accuracy (0.83%). On comparing the AUC values of the from-scratch (0.8269) and freeze models (0.5858), the performances of the other models were lower than that of the fine-tuning model. Through 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 from-scratch scheme (84.69% vs. 55.61%, p-value <0.05). In the three fine-tuning ensembled models using different data augmentation techniques, the prediction accuracy was 0.8333. Moreover, the AUC values of ADD were higher when patients with TMD were divided by age group (0.8549–0.9275) and sex (male: 0.8483, female: 0.9276). The accuracy of the ensemble model was higher than that of the human experts; however, this difference was not significant (p-value: 0.1633–0.0519). Conclusion: Our CNN model had excellent and outstanding accuracy in detecting ADD and could potentially be used by clinicians to evaluate ADD on MR images of TMD patients and improve treatment outcomes.