Hierarchical CNN-based occlusal surface morphology analysis for classifying posterior tooth type using augmented...

Area of Science:

• Dental informatics
• Computer vision
• Machine learning

Background:

• 3D digitization of dental models is increasingly used in dentistry.
• Classifying tooth types from isolated 3D point cloud models remains a significant challenge.
• Accurate tooth classification is crucial for various dental applications.

Purpose of the Study:

• To investigate an 8-class posterior tooth type classification method using convolutional neural networks (CNNs).
• To develop a hierarchical classification structure for improved tooth type identification.
• To enhance classification performance through advanced image augmentation techniques.

Main Methods:

• Utilized CNN-based occlusal surface morphology analysis for 8-class posterior tooth classification.
• Transformed 3D occlusal surfaces into depth images for CNN input.
• Implemented a two-stage hierarchical classification structure to decompose the task.
• Applied traditional geometrical transformations and deep convolutional generative adversarial networks (DCGANs) for image augmentation.

Main Results:

• Achieved an overall accuracy of 91.35% for 8-class posterior tooth type classification.
• Reported macro precision of 91.49%, macro-recall of 91.29%, and macro-F1 score of 0.9139.
• Demonstrated superior performance compared to other deep learning models.
• Grad-CAM analysis showed the CNN model focuses on smaller, important regions like anatomic landmarks (cusp, fossa, groove).

Conclusions:

• A two-stage hierarchical CNN structure effectively classifies posterior tooth types from 3D models without relative position information.
• The proposed method offers easy training and excels at learning discriminative features from limited image regions.
• This approach represents a significant advancement in automated dental model analysis.