Diagnosis of approximal caries in children with convolutional neural networks based detection algorithms on radiographs: A pilot study.

Abstract

Objectives: Approximal caries diagnosis in children is difficult, and artificial intelligence-based research in pediatric dentistry is scarce. To create a convolutional neural network (CNN)-based diagnostic system for the prompt and efficient identification of approximal caries in pediatric patients aged 5-12 years.

Materials and methods: Pediatric patients' digital periapical radiographic images were collected to create a unique dataset. Various augmentation methods were used, and approximal caries in the augmented images were labeled by a pediatric dentist to minimize labeling errors. The dataset consisted of 830 data labeled for approximal caries on 415 images, which were divided into 80% training and 20% testing sets. After comparing 13 detection algorithms, including the latest YOLOv8, the most appropriate one was selected for the proposed system, which was then evaluated based on various performance metrics.

Results: The proposed detection system achieved a precision of 91.2%, an accuracy of 90.8%, a recall of 89.3%, and an F1 value of 90.24% after 300 iterations, utilizing a learning rate of 0.01.

Conclusion: Approximal caries has been successfully detected with the developed system. Future efforts will focus on augmenting the dataset and expanding the sample size to enhance the efficacy of the system.