Related Experiment Video
Updated: Aug 17, 2025
Author Spotlight: Advancing CBCT and Digital Dental Image Integration with AI-Assisted Digitization
Published on: February 23, 2024
Feasibility of deep learning for dental caries classification in bitewing radiographs based on the ICCMS™
Wannakamon Panyarak1, Kittichai Wantanajittikul2, Wattanapong Suttapak3
1Division of Oral and Maxillofacial Radiology, Department of Oral Biology and Diagnostic Sciences, Faculty of Dentistry, Chiang Mai University, Chiang Mai, Thailand.
Objective:
To evaluate the potential of deep learning models for categorization of dental caries in bitewing radiographs based on the International Caries Classification and Management System (ICCMS™) radiographic scoring system (RSS).
Study Design:
In total, 2758 annotated bitewing radiographs were randomly divided into 3 experiments to assess the ResNet-18, -50, -101, and -152. Experiment A tested 4-class ICCMS™-RSS training and validation using Carestream (CS) radiographs; experiment B tested training and validation using CS and VistaScan radiographs; experiment C tested 7-class ICCMS™-RSS training and validation using CS and VistaScan radiographs. The performance matrices and the areas under the receiver operating characteristic curves were analyzed to assess all procedures.
Results:
In experiment A, ResNet-50 and ResNet-152 were equally accurate (71.11%) and approximately 78% sensitive. The latter presented the highest specificity (56.90%). In experiment B, ResNet-50 presented the highest sensitivity (79.51%) but ResNet-152 had the highest specificity (60.71%). In experiment C, all models markedly underperformed in distinguishing the 7-class ICCMS™-RSS with specificities of 16.46% to 22.41%. They had fewer classification errors in the 4-class classification (28.89%-35.56%) than in the 7-class classification (42.34%-53.06%). The areas under the receiver operating characteristic curves of all models were unanimously comparable.
Conclusions:
The ResNet models were able to classify dental caries according to the ICCMS™-RSS with average performances. The models underperformed in complicated classification tasks.
