Original Research Article

Fully automated deep learning framework for detection and classification of impacted mandibular third molars in panoramic radiographs

¹ Department of Oral Diagnostic Sciences, Faculty of Dentistry, SEGi University, No. 9 Jalan Teknologi, Taman Sains, Petaling Jaya, Kota Damansara, Selangor 47810, Malaysia
² School of Computing, Faculty of Computing and Engineering Technology, Asia Pacific University of Technology and Innovation (APU), Lot 6, Technology Park Malaysia, Bukit Jalil, Kuala Lumpur 57000, Malaysia
³ Oxford Internet Institute, University of Oxford, 41 St Giles, Oxford OX1 3JS, UK
⁴ Preventive Dental Science Department, College of Dentistry, King Saud Bin Abdulaziz University for Health Sciences, Riyadh 11426, Saudi Arabia
⁵ King Abdullah International Medical Research Center, Ministry of National Guard Health Affairs, Riyadh 11481, Saudi Arabia
⁶ Department of Periodontics and Implantology, Faculty of Dentistry, SEGi University, No. 9 Jalan Teknologi, Taman Sains, Petaling Jaya, Kota Damansara, Selangor 47810, Malaysia

^* Correspondence: dr.suri88@gmail.com

Received: 20 November 2024
Accepted: 3 February 2025

Abstract

Introduction: Mandibular third molars (MTMs) are the most frequently impacted teeth, making their detection and classification essential before surgical extraction. This study aims to develop and assess the accuracy of a deep learning model for detecting and classifying impacted mandibular third molars (IMTMs) using panoramic radiographs (PRs). Materials and methods: The study utilized a dataset of 1100 PRs with 1200 IMTMs and 711 PRs without MTMs. An oral radiologist validated the annotations, and the data were split into training, validation, and testing sets. The Sobel Third Molar Detection Model (STMD), built on the VGG16 architecture, identified MTMs. Detected MTMs were located using the YOLOv7 model and classified per Winter’s classification via a ResNet50-based prediction model. Results: The VGG16-based detection model achieved a testing accuracy of 93.51%, with a precision of 94.64, recall of 89.47, and an F1 score of 91.97. The ResNet50-based classification model attained a testing accuracy of 92.17%, precision of 92.1, recall of 92.17, and an AUC of 98.28. These findings demonstrate the high accuracy and reliability of both models. Conclusion: VGG16 and ResNet50 integrated with YOLOv7, demonstrated high accuracy suggesting that the automatic detection and classification of IMTMs can be significantly improved using these models.