An attention-based transfer learning model for diagnosing subluxation in temporomandibular joint panoramic radiographs.

Artificial intelligence (AI) and deep learning (DL) techniques have great potential to accelerate diagnostic processes, increase accuracy, and support clinical decision-making in healthcare. In this study, we propose a transfer learning-based approach-one of the DL techniques-to improve subluxation (SL) detection in temporomandibular joint panoramic radiography (TMJ-PR) images. For this purpose, we prepared and publicly released a dataset comprising 3,425 annotated TMJ-PR images to encourage reproducibility and further research in this domain. Several transfer learning models including MobileNet, ResNet50V2, InceptionV3, Xception, EfficientNetV2B0, InceptionResNetV2, and DenseNet201 were trained and evaluated using a 5-fold cross-validation method. By integrating a self-attention mechanism into the DenseNet201 model which achieved the highest baseline performance across all metrics, the proposed attention-based version yielded further improvements, achieving an accuracy of 90.7%, precision of 90.7%, recall of 90.7%, specificity of 89.4%, and F1-score of 90.7%. The results indicate that the proposed model achieves superior F1-score performance compared to all baseline models with relative improvements ranging from +2.40% (vs. DenseNet201) to +14.26% (vs. EfficientNetV2B0). The findings demonstrate that the proposed model not only improves subluxation detection performance but also offers a promising foundation for integration into a clinical decision support system (CDSS), enhancing early diagnosis and treatment planning using low-cost TMJ-PR images. The publicly shared dataset further supports transparency and reproducibility in future medical AI research.