Hyperparameter Tuning and Automatic Image Augmentation for Deep Learning-Based Angle Classification on Intraoral...

Area of Science:

• Artificial Intelligence in Dentistry
• Medical Image Analysis
• Orthodontic Diagnostics

Background:

• Accurate classification of dental sagittal occlusion along Angle classes is crucial for orthodontic treatment planning.
• Deep learning models offer potential for automating this classification from intraoral photographs.
• Hyperparameter tuning and data augmentation are key to optimizing deep learning performance.

Purpose of the Study:

• To evaluate the impact of hyperparameter tuning and automatic image augmentation on deep learning models for classifying orthodontic Angle classes.
• To identify optimal learning rates and batch sizes for ResNet architectures in this context.
• To assess the performance improvements gained from automatic augmentation and model explainability.

Main Methods:

• Trained ResNet architectures on a dataset of 605 Angle class I, 1038 class II, and 408 class III orthodontic images.
• Systematically tuned learning rates and batch sizes, comparing models with and without automatic image augmentation.
• Employed 10-fold cross-validation for performance evaluation and GradCAM for model explainability.

Main Results:

• Optimal hyperparameter combinations (learning rate ~1-3 × 10^-6, batch size 8) achieved accuracies of 0.63-0.64 and F1-scores of 0.61-0.62.
• Automatic image augmentation improved all performance metrics by 5-10%.
• GradCAM analysis confirmed that models utilized features relevant to human classification, with most misclassifications occurring between Angle classes I and II.

Conclusions:

• Hyperparameter selection critically influences deep learning model performance in orthodontic image classification.
• Automatic image augmentation provides substantial performance gains, enhancing diagnostic accuracy.
• The developed deep learning models demonstrate capability in classifying dental sagittal occlusion from intraoral photos.