Artificial intelligence for detecting dental ankylosis in primary molars using panoramic radiographs—a retrospective study

Background: Dental ankylosis is an eruptive abnormality that requires early diagnosis to prevent complications. This study investigated the usability and performance of various deep learning models (including transfer learning, which enhances model performance by utilizing pre-trained networks) for ankylosis detection in dental X-rays. Methods: Classical convolutional neural network (CNN) method, Visual Geometry Group 16-layer (VGG16), Inception V3, and MobileNet V2 deep learning models were used for classification. In total, 268 panoramic radiograph images were diagnosed: 98 as ankylosis cases and 170 as controls, with ages ranging from 4 to 15 years. Various data augmentation techniques were employed. Accuracy, sensitivity, specificity, Area Under Curve (AUC) and F1-Score metrics were assessed to evaluate the performance of the models. Results: The CNN network without pretraining proved insufficient, leading to the adoption of transfer learning. The accuracy, AUC, sensitivity, specificity and F1-Score values of all three models can be used, but the VGG16 and Inception V3 models generally outperformed the MobileNetV2. Based on accuracy and specificity, Inception V3 demonstrated better classification performance, while VGG16 demonstrated a more balanced performance. Conclusions: This study highlights the effectiveness of deep learning models, particularly VGG16, in identifying ankylosis from panoramic radiographs, emphasizing the importance of model selection for improved diagnostic outcomes.