Image segmentation of impacted mesiodens using deep learning.

Area of Science:

• Dentistry
• Artificial Intelligence
• Medical Imaging

Background:

• Impacted mesiodens are a common dental anomaly in children, requiring accurate and timely diagnosis.
• Early detection of mesiodens is crucial for effective orthodontic and surgical interventions.
• Traditional diagnostic methods can be time-consuming and may vary in accuracy among practitioners.

Purpose of the Study:

• To evaluate the performance of deep learning algorithms for classifying and segmenting impacted mesiodens in pediatric panoramic radiographs.
• To compare the diagnostic accuracy and efficiency of deep learning models against human expert performance.
• To assess the segmentation capabilities of a U-Net algorithm enhanced with ResNet models.

Main Methods:

• A dataset of 850 pediatric panoramic radiographs (ages 3-9) was used.
• The U-Net semantic segmentation algorithm, enhanced with pre-trained ResNet models, was employed for mesiodens detection.
• Performance metrics included Jaccard index, Dice coefficient, accuracy, precision, recall, F1-score, and diagnostic time.
• Comparisons were made against human expert diagnoses using Cohen's kappa statistics.

Main Results:

• The deep learning model achieved high segmentation performance (Jaccard index and Dice coefficient >90%).
• Diagnostic accuracy (91-92%) and F1-score (94-95%) were comparable to human experts (96%).
• The AI model diagnosed mesiodens in 7.5 seconds, significantly faster than human groups, with moderate agreement (Cohen's kappa = 0.767).

Conclusions:

• Deep learning algorithms demonstrate robust segmentation capabilities for impacted mesiodens.
• AI-powered diagnostic tools offer performance comparable to human experts in accuracy and F1-score.
• The proposed deep learning approach significantly reduces diagnostic time for mesiodens detection in pediatric dental radiography.