Second Opinion

Second Opinion® is an AI-driven software utilizing cloud-based computer vision and machine learning to analyze dental radiographs for nine types of dental features, serving as a second reader to aid dental professionals. Clinical validation, including MRMC studies with wAFROC analysis, scientifically demonstrates that the device significantly improves detection accuracy for pathologies such as caries and periapical radiolucency when used as an assistive tool by clinicians.

Second Opinion® is a computer-aided detection (CADe) software designed to assist dental health professionals in analyzing bitewing and periapical radiographs for patients aged 12 and older. It utilizes cloud-based computer vision and machine learning algorithms to identify and highlight regions suspected of containing nine dental findings, including pathologies like caries, margin discrepancies, calculus, and periapical radiolucencies, as well as restorations such as crowns, fillings, and implants, functioning as a second reader. The system architecture consists of an in-office client application, a cloud-hosted Application Programming Interface (API), and Computer Vision (CV) Models; these models process radiographic images, generate metadata detailing detected features and their properties, and the API delivers this information to the client interface for clinician review. The clinical efficacy and safety of Second Opinion® were assessed through standalone performance tests and a multi-reader, multi-case (MRMC) clinical study involving 25 readers and 2,010 images, with ground truth established by expert consensus using four expert readers. This MRMC study, employing the Weighted Alternative Free-Response Receiver Operating Characteristic (wAFROC) figure of merit, demonstrated a statistically significant improvement in detection accuracy when readers were aided by the software across all four targeted pathologic features: caries, margin discrepancy, calculus, and periapical radiolucency. Results indicated improved sensitivity for aided readers (ranging from 0.9% to 11.7% increases depending on the pathology) and, for most pathologies, a reduction in false positives per image, with no statistically significant degradation in reader performance, supporting its effectiveness as an assistive tool in dental radiographic interpretation.

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. The document highlights performance variations based on Jaccard Index (JI) thresholds (0.4 vs. 0.5); does the underlying study data, summarized here, provide a detailed rationale or sensitivity analysis for selecting JI ≥ 0.4 as optimal for both unaided reader lesion localization and device standalone performance evaluation across all detected features?; Regarding the 'open-source models' foundational to the Computer Vision algorithms, to what extent does the summarized technical documentation describe the specific models, the nature and volume of training data used for supervised machine learning, and the validation process for these adapted algorithms?; The MRMC study showed 100% of readers improved sensitivity for periapical radiolucency (PR) with aid, yet only 36% improved their false positive per image rate for PR. Does the clinical evaluation section provide further analysis on this specific outcome, such as common characteristics of PR false positives generated by the device or reader interpretations?; The device accepts multiple image formats (RVG, DICOM, JPEG, TIFF, PNG). Does the non-clinical testing summary detail the range of image acquisition equipment (e.g., different sensor brands, X-ray machine settings) from which these formats were sourced and tested to ensure consistent Computer Vision Model performance?; The document notes the software has a 'moderate level of software concern.' Does the software verification and validation testing summary elaborate on the specific risk mitigations implemented and tested, particularly concerning the cloud-based API and CV model interactions, to address potential failures or inaccuracies in a clinical setting?