Transfer learning models in the detection of pulp calcifications- A preliminary study.

Background: Pulp calcifications (PC) play a major role in the clinical outcome of endodontic treatment. Dentists use different radiographic modalities for detecting the presence of PCs. Transfer learning models have shown promising results with low computational resources in the detection and classification of diseases and conditions in dental radiographs. The aim of the present study was to evaluate the performance of transfer learning models in the detection of PCs in cropped panoramic radiographs (PRs). Materials and methods: Two calibrated examiners collected 240 cropped PR images (120 with PCs and 120 without PCs) of maxillary and/or mandibular posterior teeth. The images were preprocessed using CLAHE (Contrast Limited Adaptive Histogram Equalization) and then augmented. Three pre-trained models, VGG16, ResNet101V2, and MobileNetV2, were used for classification of the images. A Fine-tuning approach was used for training the model. Results: The inter-rater reliability among the five examiners was 0.91. VGG 16 was the best-performing model with training, validation, and test accuracies of 0.80, 0.85, and 0.85, respectively. VGG16 showed a precision of 0.84, a recall of 0.87, an F1-score of 0.86, and an AUC of 0.93. ResNet101V2 and MobileNetV2 showed test accuracies of 69% and 50%, respectively. Conclusions: The Transfer-learning model VGG 16 outperformed other models in the detection of PCs in cropped PRs. Due to the use of cropped PRs, the model cannot be generalized; however, future work will be aimed at attaining similar performance metrics in uncropped PRs in larger datasets.