Artificial intelligence-based automated preprocessing and classification of impacted maxillary canines in panoramic...

Area of Science:

• Dentistry
• Medical Imaging
• Artificial Intelligence

Background:

• Automating the digital workflow for diagnosing impacted canines using panoramic radiographs (PRs) presents significant challenges.
• Accurate identification of impacted canines is crucial for effective orthodontic and surgical treatment planning.

Purpose of the Study:

• To explore feature extraction, automated cropping, and classification of impacted and nonimpacted canines using AI as a preliminary step.
• To enhance the accuracy of impacted canine diagnosis through automated preprocessing of PRs.

Main Methods:

• A convolutional neural network (CNN) with SqueezeNet architecture was trained for canine impaction classification.
• AI detectors were developed to identify key anatomical landmarks on PRs for automated cropping.
• The impact of automated cropping guided by landmarks on classification accuracy was evaluated.

Main Results:

• Initial classification accuracy without cropping achieved an area under the curve (AUC) of 84% on the receiver operating characteristic (ROC) curve.
• AI landmark detectors accurately identified landmarks in approximately 98% of PRs.
• Automated cropping using specific landmarks improved the AUC-ROC to 96% for impacted canine classification.

Conclusions:

• AI algorithms can be effectively automated to preprocess PRs.
• Automated preprocessing significantly enhances the identification accuracy of impacted canines.