Can Artificial Intelligence help predict the risk of inferior alveolar nerve injury before mandibular third-molar surgery?

1. Title

A Fused Deep Learning Architecture for the Detection of the Relationship between the Mandibular Third Molar and the Mandibular Canal

DOI

https://doi.org/10.3390/diagnostics12082018

2. Authors

Cansu Buyuk et. al

3. Summary

This study evaluated a fused deep learning architecture combining a U-Net model for segmentation and an AlexNet model for classification to assess the anatomical relationship between the mandibular third molar (M3) and the mandibular canal (MC) on orthopantomographs (OPGs). Using 1,880 panoramic radiographs, the U-Net network segmented the regions of interest (achieving 0.99 global accuracy and 0.91 mean Dice coefficient), while AlexNet classified the degree of overlap into four distinct categories (achieving 0.80 accuracy and 0.85 AUC). The performance of the deep learning model closely matched that of experienced dental practitioners (0.79 success rate), indicating its potential as an effective clinical decision support tool for pre-surgical risk assessment and minimizing inferior alveolar nerve injury.

4. Key Words

deep learning; segmentation; third molar; mandibular canal; panoramic radiography

5. Extracted Data

5.1 Year

2022

5.2 Modality

Orthopantomogram (OPG) / Panoramic Radiography

5.3 Dataset

1,880 OPGs (2,893 image patches / regions of interest). Number of unique patients: Not reported.

5.4 Dataset Split

80% Training / 20% Testing (U-Net: 1,504 training / 376 testing; AlexNet: 2,348 training patches / 545 testing patches)

5.5 Network Architecture

Fused Architecture: U-Net (for ROI segmentation) + AlexNet (for ROI classification)

5.6 Metrics

• Segmentation (U-Net): Global Accuracy = 0.99, Weighted IoU = 0.98, Mean Dice Coefficient = 0.91
• Classification (AlexNet): Accuracy = 0.80, AUC = 0.85

5.7 Ground-truth - Professional Qty

2 Dentomaxillofacial Radiologists

5.8 Ground-truth - Supervisor Presence

Not reported

5.9 Ground-truth - Experience Level

Not reported

5.10 Ground-truth - Expertise Area

Dentomaxillofacial Radiology

5.11 Ground-truth - Tool or System

MATLAB (The MathWorks, Inc., Natick, MA, USA)

5.12 ML Task

Combined Semantic Segmentation and Multi-class Classification (4 overlap classes)

5.13 Project Objective

To generate a fused deep learning algorithm that automatically detects and classifies the relationship and degree of overlap between the mandibular third molar and the mandibular canal on panoramic radiographs to assist in surgical risk prediction.

6. Clinical Relevance

Extraction of mandibular third molars is a common dental procedure where inferior alveolar nerve (IAN) injury is a major risk. Automated, highly accurate AI risk-assessment on routine panoramic radiographs provides clinicians with an accessible decision-support tool to evaluate anatomical proximity, optimize surgical planning, and minimize nerve damage without requiring routine high-radiation 3D CBCT scans.