Deep learning-based detection of incisal translucency patterns.

Area of Science:

• Artificial Intelligence in Dentistry
• Computer Vision for Dental Diagnostics
• Digital Dentistry

Background:

• Incisal translucency evaluation is crucial for dental esthetics but is often subjective.
• Current methods for assessing translucency are inconsistent and overlooked by professionals.
• Objective, AI-driven tools are needed to support dentists in restorative procedures.

Purpose of the Study:

• To evaluate the accuracy of deep learning models in predicting anterior tooth translucency patterns.
• To develop an AI system for objective assessment of incisal translucency.
• To provide a quantitative tool for dentists to aid in esthetic treatment planning.

Main Methods:

• A dataset of 240 anterior tooth JPEG images was collected via smartphone.
• A 3-model deep learning pipeline was employed: YOLOv5 for detection, Vision Transformers (ViT) for classification, and U-Net for segmentation.
• Image augmentation and an 80/20 train-test split were used, with performance metrics including accuracy, precision, recall, F1 score, and dice scores.

Main Results:

• YOLOv5 achieved perfect precision (1.00) at a 0.910 confidence threshold for detecting central incisors.
• The Vision Transformers (ViT) model demonstrated 91.66% accuracy and a 94.83% F1 score for translucency identification.
• U-Net segmentation achieved 91% accuracy and a 0.948 dice score for outlining translucent areas.

Conclusions:

• A combined deep learning approach using YOLOv5, ViT, and U-Net effectively classifies incisal translucencies.
• This AI system offers high accuracy and precision for detecting anterior tooth translucency patterns.
• The developed models provide a valuable, objective tool for dentists in esthetic restorative dentistry.