Velmeni for Dentists (V4D)

VELMENI for DENTISTS (V4D) is a machine learning-driven software intended to support dentists by concurrently analyzing digital bitewing, periapical, and panoramic radiographs to identify dental caries, restorations, fixed prostheses, and implants in patients aged 15 and older. Its system architecture incorporates a web interface and an ML engine for image classification and condition detection, with performance demonstrated through standalone tests and clinical reader studies showing improvements in detection accuracy for various dental conditions and radiograph types.

Velmeni for Dentists (V4D) is a computer-assisted detection software developed to aid dental professionals in the clinical identification of dental caries, fillings/restorations, fixed prostheses, and implants. The system is designed to analyze digital bitewing, periapical, and panoramic radiographs from patients aged 15 years or older who possess permanent dentition. Functioning as a concurrent-read tool, V4D provides supplementary information to the examining dentist and is not intended as a substitute for their comprehensive clinical judgment, which considers all patient information. The core of V4D is its Machine Learning (ML) Engine, which processes grayscale radiographs (jpeg and png formats) using several modules: a radiograph type classifier, a condition detection module, a tooth numbering module, and a merging module to consolidate findings. The overall system architecture includes a web application interface for user interaction, a backend API for inter-component communication, a queue for managing analysis requests, an AI-worker to execute ML tasks, and dedicated database and file storage for critical information. The performance of V4D was rigorously evaluated through both standalone testing and a clinical reader study. Standalone tests were conducted on a dataset comprising 600 bitewing, 597 periapical, and 600 panoramic images, measuring lesion-level sensitivity for caries (e.g., 72.8% for bitewing views) and other metrics such as sensitivity, specificity, and dice coefficients for fixed prostheses, implants, and restorations across the different radiograph types. For example, lesion-level sensitivity for fixed prostheses in bitewing views reached 92.1%. The clinical reader study involved 12 U.S. licensed dentists who evaluated images with and without the assistance of V4D. This study demonstrated improved weighted Alternative Free-response Receiver Operating Characteristic (wAFROC) Area Under the Curve (AUC) for caries detection across all views when aided by V4D (e.g., bitewing wAFROC AUC increased from 0.794 unaided to 0.848 aided). For bitewing images, V4D assistance notably increased lesion-level sensitivity for caries from 67.5% to 80.3% and for restorations from 74.1% to 90.8%, with minimal change in false positives per image for caries. Similar improvements were observed for periapical views, although there was a decrease in case-level specificity for caries. For panoramic views, V4D provided a benefit, particularly increasing restoration sensitivity, though this was accompanied by a reduction in image-level specificity for restorations.

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 algorithm differentiate between primary and secondary caries, and what were the specific performance metrics (e.g., sensitivity, false positives per image) for each of these subgroups in the standalone and clinical studies, given that subgroup analysis was performed?; The document lists Dexis, Kavo, Carestream, and Planmeca as sensor manufacturers for standalone testing. Were other sensor types or models beyond these specific ones evaluated, and did the device's performance show significant variability across different sensor technologies that could impact its generalizability in diverse clinical settings?; The 'Rejection Review (RR) module' is designed to capture ML-detected conditions rejected by dental professionals. Can you elaborate on the operational workflow of this module and detail how the collected data on rejected detections is systematically analyzed and incorporated into the feedback loop for ML model refinement and future product iterations?; For panoramic views in the clinical study, it was noted that the benefit in restoration sensitivity was 'somewhat offset by a drop in image-level specificity.' What were the precise values for this trade-off (i.e., the magnitude of sensitivity gain versus specificity loss), and what are the potential underlying factors or algorithmic characteristics contributing to the increased false positives for restorations specifically in panoramic images when clinicians are aided by the software?; The 'tooth numbering module' is a key component for localizing findings. What methodologies does this module employ to accurately identify and number teeth, particularly when encountering challenging dental anatomies such as missing teeth, supernumerary teeth, significant malocclusion, or unerupted teeth, and what was its standalone performance accuracy in these scenarios?