DentalCLIP: Semi-Supervised Contrastive Learning for TMJ Osteoarthritis Diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Osteoarthritis Research

Background:

• Temporomandibular Joint Osteoarthritis (TMJOA) is a common degenerative condition causing pain and functional issues.
• Early diagnosis is crucial but challenging due to subtle early changes.
• Current methods require multiple imaging types, limiting data availability.

Purpose of the Study:

• To develop an effective diagnostic tool for TMJOA using only panoramic radiography (PR).
• To leverage contrastive semi-supervised learning to overcome limitations of scarce labeled medical data.
• To improve the accuracy and accessibility of TMJOA diagnosis.

Main Methods:

• Proposed DentalCLIP, a framework fine-tuning CLIP using contrastive learning and pseudo-labeling.
• Utilized panoramic radiography (PR) for TMJOA detection.
• Evaluated performance against vanilla CLIP, BiomedCLIP, and MedCLIP.

Main Results:

• DentalCLIP achieved high diagnostic accuracy (0.929 ± 0.012) and F1-score (0.929 ± 0.012).
• Outperformed vanilla CLIP and domain-specific models (BiomedCLIP, MedCLIP).
• Demonstrated the efficacy of contrastive semi-supervised learning in medical imaging.

Conclusions:

• DentalCLIP offers a promising approach for enhanced TMJOA diagnosis using PR.
• The framework effectively utilizes unlabeled data, addressing data scarcity issues.
• Highlights the potential of AI in improving diagnostic performance in resource-limited settings.