Comprehensive Summary
This study aims to analyze the ability of various convolutional neural networks (CNNs) to analyze and interpret panoramic radiographs (PRs), thereby better informing clinical decisions regarding impacted maxillary canines (IMCs). Patient data of PRs with late mixed dentition, complete permanent dentition, no technical or patient-related artifacts, and confirmed diagnosis of IMCs or non-IMCs were collected for training. PRs were graded by two orthodontists, a resident and a supervising faculty member with 15+ years of experience. Sensitivity, specificity, accuracy, and Area Under the Curve of the Receiver Operating Characteristic (AUC-ROC, diagnostic accuracy metric) were evaluated for all CNNs. With respect to AUC-ROC, the CNN GoogLeNet performed the best, and was even able to detect and classify images of fractured dental implants. Overall, each CNN performed relatively well, with each one excelling at distinct tasks.
Outcomes and Implications
This study presents a well-researched and comprehensive overview of various commonly used CNNs and their capabilities in clinical examination of PRs for IMCs. Given more time for the implementation strategy and development of more generalizable models, there is a high possibility that clinical implementation of CNNs in orthodontic assessment can begin soon. Clinicians often have many imaging techniques that require careful analysis for the development of treatment plans and assessment of conditions. This study offers the abilities of CNNs as a supplement to aid clinicians in the diagnosis and treatment of maxillary canine impaction.