Overjet Caries Assist

Overjet Caries Assist is an artificial intelligence-powered software designed to aid dentists by improving the detection and segmentation of carious lesions on adult bitewing radiographs. Utilizing an ensemble of U-Net based machine learning models, the system processes images to provide annotated suggestions, which clinical studies have shown to significantly enhance dentists' diagnostic sensitivity for caries.

Overjet Caries Assist (OCA) is a software-based medical image analyzer designed to assist dentists in detecting and segmenting carious lesions on bitewing radiographs. It is intended for adult patients (18 years and older) with permanent dentition and functions as an adjunctive tool, providing additional information rather than replacing a clinician's comprehensive review and judgment. The system operates using a three-layer architecture (Network, Presentation, Decision) and employs a machine learning model within its Decision Layer which processes bitewing radiographs through several modules: an Image Classifier to validate image type, a Tooth Number Assignment module for pixel-wise tooth segmentation, a Caries Module utilizing an ensemble of three U-Net based models for pixel-wise caries segmentation, and a Post Processing module to assign detected caries to specific teeth and annotate the original radiograph. Performance testing of Overjet Caries Assist included standalone algorithm evaluation and a clinical reader study. In standalone tests, the algorithm demonstrated an overall sensitivity of 72.0% and specificity of 98.1% for caries detection on tooth surfaces, with mean Dice scores for lesion segmentation ranging from 0.69 to 0.75 for primary and secondary caries respectively. The clinical reader study, involving 13 dentists, showed that using OCA significantly improved their average sensitivity in detecting caries from 57.9% to 76.2%, while specificity slightly decreased from 99.3% to 98.4%; furthermore, the average area under the Alternative Free-Response Receiver Operating Characteristic (AFROC) curve increased by 0.057 (p < 0.001), indicating enhanced diagnostic performance. Key limitations include its inability to detect gross decay or caries obscured by restorations, and the potential for false positives/negatives, emphasizing the need for final clinical judgment.

AI Regulatory Concerns

Powered by

Jev reads the full FDA letter and returns calibrated concern probabilities per regulatory category. These are AI estimates, not regulatory advice.

• Q1. How does the performance of the ensemble of three U-Net based models in the Caries Module compare to individual U-Net models or other deep learning architectures for caries segmentation accuracy and robustness?; What specific image characteristics (e.g., sensor type, exposure settings, presence of artifacts) most significantly impact the standalone performance (sensitivity and specificity) of the Overjet Caries Assist algorithm?; The study reported a slight decrease in specificity (0.9%) when clinicians used the OCA. What are the potential reasons for this decrease, and what steps are being taken to mitigate false positives without compromising sensitivity gains?; Given that the device is not intended for gross decay, how does the algorithm differentiate between incipient/moderate caries (which it is trained to detect) and gross decay, and what is the threshold for a lesion to be considered "gross decay" and thus outside the system's supported detection capabilities?; The Dice scores for segmentation improved with OCA assistance but were not always statistically significant in the reader study. What further developments are planned to enhance the precision of automated lesion boundary delineation, and how critical is precise segmentation to the clinical utility beyond simple detection?