The validity of an artificial intelligence application for assessment of orthodontic treatment need from clinical images
MetadataShow full item record
Aim: To assess the validity of a Convolutional Neural Network (CNN) digital model to detect and localize orthodontic malocclusions from intraoral clinical images. Materials and methods: The sample of this study consisted of the intraoral images of 700 Subjects. All images were intraoral clinical images, in one of the following views: Left Occlusion, Right Occlusion, Front Occlusion, Upper Occlusal, and Lower Occlusal. The following malocclusion conditions were localized: crowding, spacing, increased overjet, cross bite, open bite, deep bite. The images annotations were repeated by the same investigator (S.T) with a one week interval (ICC 0.9). The CNN model used for this research study was the “You Only Look Once” model. This model can detect and localize multiple objects or multiple instances of the same object in each image. It is a fully convolutional deep neural network; 24 convolutional layers followed by 2 fully connected layers. This model was implemented using the TensorFlow framework freely available from Google. Results: The created CNN model was able to detect and localize the malocclusions with an accuracy of 99.99%, precision of 99.79%, and a recall of 100%. Conclusions: The use of computational deep convolutional neural networks to identify and localize orthodontic problems from clinical images proved valid. The built AI engine accurately detected and localized malocclusion from different views of intra-oral clinical images.