Dental Anomaly Image Classification Using a Convolutional Neural Network

Objectives: Children with orofacial clefting (OFC) present with a wide range of complex dental anomalies. Identifying these anomalies is vital to understanding the complex phenotypic spectrum of OFC. Such anomalies are currently identified using intra-oral exams and/or photos and this process is time consuming, particularly for large samples. Therefore, automating the process of anomaly detection using neural networks (NN) could increase reliability and speed of anomaly detection. This study characterizes the use of NN to identify dental anomalies using intraoral photographs in one of the largest international cohorts to date of children with nonsyndromic OFC and controls.
Methods: Intraoral images from 1,937 subjects were previously scored for dental anomalies by human raters whom achieved inter-rater reliability of kappa=0.91-0.93. A convolutional neural network (CNN) performed multi-class classification of 8 dental anomalies (mammelons, impacted, hypoplasia, rotated, incisal fissures, fluorosis, hypocalcification, and displaced) using a photo for each subject simultaneously displaying the maxillary and mandibular canine-to-canine region. The network predicts whether a patient exhibits any of the aforementioned anomalies. Training and testing of the CNN on the total sample were performed in an 80/20% split respectively. The networks architecture consisted of two convolutional and two linear layers. False positive (FPR) and false negative (FNR) rates were determined for each anomaly.
Results: The CNN, for all subjects, identified the dental anomalies, in microseconds per photo, with an average accuracy of 68.20%. Per anomaly, accuracy ranged from 50-100%, FPR from 0-8% and FNR from 0-46%, except for rotation and hypocalcification.
Conclusions: We show that a naïve neural network image classifier was able to attain an accuracy consistent with an interrater-reliability kappa of substantial agreement, suggesting that the CNN has the potential to be an alternative rating method of dental anomalies in intraoral photographs, thus decreasing time spent and facilitating dental anomaly rating standardization for future large multicenter studies.