Artificial Intelligence-Assisted Orthodontic Extraction Decision Support Using Routine Intraoral Photographs: A Clinical Benchmark of Deep Learning Architectures

Background Orthodontic extraction decision-making is a complex process requiring comprehensive clinical evaluation. Although artificial intelligence (AI) has shown promise in supporting orthodontic diagnosis and treatment planning, the potential of routinely acquired intraoral photographs as the sole imaging source for AI-assisted extraction decision support remains unclear. This study aimed to evaluate the feasibility of image-based AI for orthodontic extraction decision support and to benchmark the performance of multiple deep learning architectures using routine intraoral photographs. Methods A total of 2500 patients with Class I malocclusion, each with 5 standardised pretreatment intraoral photographs, were categorised as extraction or nonextraction based on completed clinical treatment records. Eleven convolutional neural network architectures were trained using transfer learning to predict extraction decisions from the right lateral intraoral photograph of each patient. The architecture with the highest observed accuracy, VGG11, was subsequently evaluated in a separate multi-image experiment using all 5 intraoral views. Model performance was assessed using accuracy, precision, recall, F1-score and Gradient-weighted Class Activation Mapping (Grad-CAM). Results Among the evaluated architectures, VGG11 achieved the highest observed classification accuracy (75.6%), while VGG13 demonstrated the highest recall (88.0%) and F1-score (77.7%) for extraction cases. In the external validation cohort, VGG11 achieved an accuracy of 73.3% (95% CI, 60.3%-83.9%), with a precision of 67.5%, recall of 90.0%, and F1-score of 77.1%. Grad-CAM visualisations showed that model activation was predominantly observed in regions containing dental alignment and occlusal structures. In the separate multi-image experiment, VGG11 showed limited classification performance (accuracy, 50.0%) with a strong tendency to predict extraction cases. Conclusion Routine intraoral photographs contain clinically meaningful information for predicting orthodontists’ extraction decisions. Among the evaluated architectures, VGG11 achieved the highest observed classification accuracy. These findings provide proof-of-concept evidence for image-based AI as a preliminary adjunct to orthodontic extraction decision-making, while emphasising that comprehensive clinical evaluation remains essential.

Authors

Institutions

Publication Details

Journal
International Dental Journal
Published
2026-11-13
DOI
https://doi.org/10.1016/j.identj.2026.111311
Primary Topic
Orthodontics and Dentofacial Orthopedics
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Artificial Intelligence-Assisted Orthodontic Extraction Decision Support Using Routine Intraoral Photographs: A Clinical Benchmark of Deep Learning Architectures

Kyungmin Clara Lee
3 links
International Dental Journal
Orthodontics and Dentofacial Orthopedics
article

Artificial Intelligence-Assisted Orthodontic Extraction Decision Support Using Routine Intraoral Photographs: A Clinical Benchmark of Deep Learning Architectures

Kyungmin Clara Lee
article en

Abstract

Background Orthodontic extraction decision-making is a complex process requiring comprehensive clinical evaluation. Although artificial intelligence (AI) has shown promise in supporting orthodontic diagnosis and treatment planning, the potential of routinely acquired intraoral photographs as the sole imaging source for AI-assisted extraction decision support remains unclear. This study aimed to evaluate the feasibility of image-based AI for orthodontic extraction decision support and to benchmark the performance of multiple deep learning architectures using routine intraoral photographs. Methods A total of 2500 patients with Class I malocclusion, each with 5 standardised pretreatment intraoral photographs, were categorised as extraction or nonextraction based on completed clinical treatment records. Eleven convolutional neural network architectures were trained using transfer learning to predict extraction decisions from the right lateral intraoral photograph of each patient. The architecture with the highest observed accuracy, VGG11, was subsequently evaluated in a separate multi-image experiment using all 5 intraoral views. Model performance was assessed using accuracy, precision, recall, F1-score and Gradient-weighted Class Activation Mapping (Grad-CAM). Results Among the evaluated architectures, VGG11 achieved the highest observed classification accuracy (75.6%), while VGG13 demonstrated the highest recall (88.0%) and F1-score (77.7%) for extraction cases. In the external validation cohort, VGG11 achieved an accuracy of 73.3% (95% CI, 60.3%-83.9%), with a precision of 67.5%, recall of 90.0%, and F1-score of 77.1%. Grad-CAM visualisations showed that model activation was predominantly observed in regions containing dental alignment and occlusal structures. In the separate multi-image experiment, VGG11 showed limited classification performance (accuracy, 50.0%) with a strong tendency to predict extraction cases. Conclusion Routine intraoral photographs contain clinically meaningful information for predicting orthodontists’ extraction decisions. Among the evaluated architectures, VGG11 achieved the highest observed classification accuracy. These findings provide proof-of-concept evidence for image-based AI as a preliminary adjunct to orthodontic extraction decision-making, while emphasising that comprehensive clinical evaluation remains essential.

International Dental JournalVol. 77(1)
Chonnam National University (KR)
National Research Foundation of Korea, Ministry of Science and ICT, South Korea
Orthodontics and Dentofacial Orthopedics
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon