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

Area of Science:

• Artificial Intelligence in Medical Imaging
• Dental Radiology
• Deep Learning Applications

Background:

• Panoramic radiographs (PRs) are crucial for dental diagnostics.
• Head positioning errors can compromise PR image quality and diagnostic accuracy.
• Automated detection of these errors is needed for standardization.

Purpose of the Study:

• To evaluate the effectiveness of deep learning (DL) architectures for automatic classification of head positioning errors in PRs.
• To compare the performance of convolutional neural network (CNN) and vision transformer (ViT) models.
• To identify optimal DL models for improving PR quality.

Main Methods:

• Retrospective collection of 480 anonymized PR images.
• Dataset categorized by an experienced oral radiologist for occlusal plane orientation.
• Splitting data into training (70%), validation (15%), and testing (15%) sets.
• Fine-tuning pre-trained CNN and ViT models using transfer learning.
• Performance evaluation using accuracy, precision, recall, F1-score, and ROC-AUC.

Main Results:

• ResNet18 (a CNN) achieved the highest performance with an accuracy and F1-score of 0.84.
• Vision transformer (ViT) models showed lower performance (accuracy and F1-score: 0.77), potentially due to smaller dataset size.
• ResNet18 demonstrated superior discriminative capability compared to ViT, indicated by ROC-AUC values.

Conclusions:

• This study is the first to comprehensively evaluate both CNN and ViT architectures for detecting occlusal plane positioning errors in PRs.
• CNN-based models, specifically ResNet18, are effective in identifying these errors.
• Implementing these DL models can reduce positioning mistakes, enhance image quality, and improve diagnostic reliability in dental practice, especially for less experienced operators.