An Artificial Intelligence Solution for Automated Dental Inspection and Charting

Visual inspection and documentation (charting) of findings are critical components in dental practice. However, manual charting is tedious, time consuming, and prone to errors. Attempts to automate charting using AI have focused on computer vision of radiographs, which does not replace visual inspection and has limitations in community screening settings. We aim to develop a deep-learning-based real-time charting using intra-oral videos. This study encompasses tooth identification, the initial phase of our development process.

100 intra-oral videos were recorded, annotated using CVAT software, and split into training and testing datasets (80:20 ratio) to develop an object detection YOLOv7x model on a Convolutional Neural Network (CNN). The model was evaluated by its ability to accurately detect each tooth's ID (class) and position using the mean average precision (AP), which encases a confusion matrix, Precision and Recall, based on a minimum overlap of 50% between manual and automated annotation (Intersection over Union IoU = 0.5).

Of the 186,332 captured image frames; 8747 were annotated, with data augmentation yielding 17,494 images. AP varied among tooth types; mAP at 0.5 IoU reached 78.8%, with Precision and Recall at 72.5% and 77.9% respectively.

To the best of our knowledge, this is the first study to utilize real-time videos for this purpose. This research presents a novel approach to automate dental inspection and charting, offering the potential to reduce clinical time and costs, improve patient experience and outcomes, and enhance patient access to dental care.