Deep learning for osteoarthritis classification in temporomandibular joint

Area of Science:

• Medical imaging analysis
• Artificial intelligence in diagnostics
• Oral and maxillofacial radiology

Background:

• Temporomandibular joint osteoarthritis (TMJ-OA) diagnosis relies on imaging.
• Developing accurate and efficient diagnostic tools for TMJ-OA is crucial.
• Panoramic radiography is a common imaging modality for TMJ assessment.

Purpose of the Study:

• To develop a diagnostic support tool for classifying temporomandibular joint (TMJ) panoramic images.
• To differentiate between normal TMJ and TMJ osteoarthritis (OA) using pretrained deep learning models.

Main Methods:

• Utilized 858 TMJ panoramic images (395 normal, 463 TMJ-OA) from 518 individuals.
• Employed transfer learning with pretrained Resnet152 and EfficientNet-B7 models.
• Evaluated model performance using accuracy, specificity, sensitivity, AUC, and grad-CAM, comparing with dentists.

Main Results:

• ResNet-152 achieved 0.87 accuracy, and EfficientNet-B7 achieved 0.88 accuracy.
• Both models demonstrated high accuracy in classifying TMJ-OA.
• Gradient-weighted class activation mapping (grad-CAM) analysis showed models focused on pathological areas like erosion and osteophytes.

Conclusions:

• AI models trained on 2D panoramic condyle images enhance diagnostic capabilities for TMJ-OA.
• The developed AI tool can serve as an effective screening diagnostic aid for dentists.
• Deep learning models show significant potential in improving TMJ-OA detection accuracy.