Intraoral radiograph anatomical region classification using neural networks

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Computer Vision for Medical Diagnostics

Background:

• Dental radiography constitutes 13% of all radiological diagnostic imaging procedures.
• Manual classification of intraoral radiographs is time-consuming and can impact metadata quality.
• Automating this classification is challenging due to limited anatomical variation and potential image overlap.

Purpose of the Study:

• To investigate the feasibility of using neural networks for automated classification of anatomical regions in intraoral radiographs.
• To develop and evaluate 36 neural network models for classifying 22 unique anatomical classes.
• To assess the impact of different data augmentation strategies and model architectures on classification performance.

Main Methods:

• Systematic development and training of 36 neural network models using full supervision and three data augmentation strategies.
• Utilized libre software and limited computational resources for model development.
• Trained and validated models on a dataset of 15,254 intraoral radiographs, evaluating performance using Top-1 accuracy, AUC, and F1-score.

Main Results:

• Cochran's Q test revealed statistically significant differences in classification performance across all models (p < 0.001).
• Advanced deep learning architectures (VGG16, MobilenetV2, InceptionResnetV2) demonstrated greater robustness to image distortions compared to baseline models.
• Top-performing models achieved classification accuracy between 81-89%, F1-scores of 0.71-0.86, and AUC values of 0.86-0.94.

Conclusions:

• Automated classification of anatomical classes in digital intraoral radiographs is achievable with high accuracy (nearly 90%), even with distorted or overlapping anatomy.
• Key factors influencing classification performance include model architecture, data augmentation, pooling/normalization layers, and model capacity.
• This automation holds significant potential for time savings and enhanced metadata quality in dental radiology.