IADR Abstract Archives
Deep Learning-based Diagnosis of Temporomandibular Joint Osteoarthritis Using Radiological Images
Objectives:
Temporomandibular joint (TMJ) osteoarthritis is a common chronic degenerative joint disease. Currently, TMJ osteoarthritis is diagnosed by symptom assessment and evaluation of cone-beam computed tomography (CBCT) findings. However, this process is subjective. In this study, we present a novel, automated method based on the Deep Siamese Convolutional Neural Network to diagnose cases of TMJ osteoarthritis using a binary classifier.
Methods:
CBCT data obtained from the Kyung Hee University Dental Hospital TMJ Osteoarthritis Study was used to train the deep learning algorithm, and randomly selected 1,000 subjects (2000 TMJ cases) from the Osteoarthritis Initiative dataset, collected and established from January 1, 2015, to September 1, 2019, was used for validating the algorithm.
Results: Our method yielded a quadratic Kappa coefficient of 0.87 and an average multiclass accuracy of 68.5% as compared to the annotations provided by a clinical experts committee which diagnosed the TMJ osteoarthritis cases according to the Research Diagnostic Criteria for Temporomandibular Disorders Axis I. Additionally, the area under the receiver operating characteristic curve for the radiological diagnosis of TMJ osteoarthritis with our algorithm was 0.91. Furthermore, combination of magnetic resonance imaging with CBCT remarkably improves the prediction accuracy as indicated by a quadratic Kappa coefficient of 0.97. Moreover, we generated attention maps to highlight the radiological features that affect the network decision.
Conclusions:
Our results ease the decision-making process clear for clinicians and researchers, which in turn, improves the reliability of artificial intelligence-based diagnostic algorithms. We believe that our model is useful for clinical decision-making and for research on TMJ osteoarthritis.
Temporomandibular joint (TMJ) osteoarthritis is a common chronic degenerative joint disease. Currently, TMJ osteoarthritis is diagnosed by symptom assessment and evaluation of cone-beam computed tomography (CBCT) findings. However, this process is subjective. In this study, we present a novel, automated method based on the Deep Siamese Convolutional Neural Network to diagnose cases of TMJ osteoarthritis using a binary classifier.
Methods:
CBCT data obtained from the Kyung Hee University Dental Hospital TMJ Osteoarthritis Study was used to train the deep learning algorithm, and randomly selected 1,000 subjects (2000 TMJ cases) from the Osteoarthritis Initiative dataset, collected and established from January 1, 2015, to September 1, 2019, was used for validating the algorithm.
Results: Our method yielded a quadratic Kappa coefficient of 0.87 and an average multiclass accuracy of 68.5% as compared to the annotations provided by a clinical experts committee which diagnosed the TMJ osteoarthritis cases according to the Research Diagnostic Criteria for Temporomandibular Disorders Axis I. Additionally, the area under the receiver operating characteristic curve for the radiological diagnosis of TMJ osteoarthritis with our algorithm was 0.91. Furthermore, combination of magnetic resonance imaging with CBCT remarkably improves the prediction accuracy as indicated by a quadratic Kappa coefficient of 0.97. Moreover, we generated attention maps to highlight the radiological features that affect the network decision.
Conclusions:
Our results ease the decision-making process clear for clinicians and researchers, which in turn, improves the reliability of artificial intelligence-based diagnostic algorithms. We believe that our model is useful for clinical decision-making and for research on TMJ osteoarthritis.
