Towards accurate occlusal plane positioning in panoramic radiographs: a deep learning-assisted study

To assess the effectiveness of deep learning architectures in automatically classifying head positioning errors on panoramic radiographs (PRs).

A total of 480 anonymized PR images were retrospectively collected and categorized by an experienced oral radiologist based on occlusal plane orientation. The dataset was randomly split into 70% for training, 15% for validation, and 15% for testing to ensure balanced model evaluation. Several pre-trained convolutional neural network (CNN) and vision transformer (ViT) models were fine-tuned using transfer learning (TL) strategy. Models were evaluated using widely adopted performance metrics including accuracy, precision, recall, F1-score, and area under curve based on receiver operating characteristic (ROC-AUC).

Among the tested architectures, ResNet18 achieved the best performance, with an overall test accuracy and F1-score of 0.84. The ViT model demonstrated comparatively lower performance (accuracy and F1-score: 0.77), which may be attributable to the relatively small dataset size. In terms of ROC-AUC performance indicator, ResNet18 also outperformed the ViT model across all classes, revealing superior discriminative capability.

To the best of our knowledge, no prior study has comprehensively evaluated both convolution-based and transformer-based deep learning architectures for detecting occlusal plane positioning errors in PRs. The findings suggest that CNN-based models-particularly ResNet18-can effectively identify such errors. These models may help reduce positioning mistakes, especially in PRs acquired by students or less experienced operators, thereby promoting increased standardization, improved image quality, and enhanced diagnostic reliability in dental practice.