DentalCLIP: Semi-Supervised Contrastive Learning for TMJ Osteoarthritis Diagnosis

Temporomandibular Joint Osteoarthritis (TMJOA) is a prevalent degenerative disorder characterized by pain, functional impairment, and malocclusion, profoundly impacting patients' quality of life. Early and precise diagnosis is critical to mitigate disease progression; however, the detection of subtle morphological changes in the initial stages presents significant diagnostic challenges. Conventional diagnosis necessitates both cone-beam computed tomography (CBCT) and panoramic radiography (PR) from the same patient, limiting the availability of comprehensive and well-annotated datasets. To address this limitation, we propose DentalCLIP, a contrastive semi-supervised learning framework designed to enhance TMJOA diagnosis using PR. DentalCLIP fine-tunes the contrastive language-image pretraining (CLIP) model by aligning image and text features through contrastive learning, while incorporating pseudo-labeling to leverage unlabeled data, thereby mitigating the scarcity of labeled medical datasets. Experimental results demonstrate that DentalCLIP not only outperforms the vanilla CLIP model but also surpasses domain-specific pretrained models such as BiomedCLIP and MedCLIP on our dataset, attaining an accuracy of 0.929 ± 0.012 and an F1-score of 0.929 ± 0.012. These findings underscore the effectiveness of contrastive semi-supervised learning in medical imaging, enhancing diagnostic performance and accessibility, particularly in resource-constrained scenarios with limited labeled data.