Rayvolve

Rayvolve is a deep learning-based software designed to aid clinicians in detecting musculoskeletal fractures in adult and pediatric patients (≥2 years) by analyzing DICOM-standard radiographs across thirteen specified anatomical regions and their associated views. Performance evaluations, including standalone testing on distinct adult and pediatric datasets as well as a multi-reader multi-case clinical study, demonstrate its efficacy in enhancing diagnostic accuracy, sensitivity, and specificity for fracture identification when used as an assistive diagnostic tool.

Rayvolve is a standalone software medical device that employs deep learning algorithms to assist radiologists and emergency physicians in the detection and localization of fractures on osteoarticular X-rays. Designed for both adult and pediatric patients (aged 2 years and older), Rayvolve processes DICOM standard medical images and can integrate with existing Picture Archiving and Communication Systems (PACS). It provides aided diagnosis for a comprehensive range of musculoskeletal radiographs, covering anatomic areas such as the ankle, clavicle, elbow, forearm, hip, humerus, knee, pelvis, shoulder, tibia/fibula, wrist, hand, and foot, across various industry-standard radiographic views. The software is intended to be used as an adjunctive tool, with its output appearing after the initial image, and does not operate autonomously; users are trained via specific instructions for use. The performance of Rayvolve has been assessed through several studies detailed in its documentation. A standalone performance assessment on 3016 pediatric radiographs (≥2 years) demonstrated high sensitivity (0.9611), specificity (0.8597), and an Area Under the Curve (AUC) of 0.9399 for fracture detection. The predicate adult version, tested on 2626 radiographs, showed high sensitivity (0.98763), specificity (0.88558), and an AUC of 0.98607. An updated algorithm, retrained with an expanded dataset of 150,000 radiographs (compared to 115,000 previously) to include a more comprehensive representation of pediatric cases alongside adult cases, demonstrated non-inferior performance to the predicate on adult data, achieving an overall AUC of 0.98781. Furthermore, a multi-reader multi-case (MRMC) clinical study involving 24 readers and 186 cases found that Rayvolve significantly improved reader diagnostic accuracy: reader AUC increased from 0.84602 (unaided) to 0.89327 (aided), mean sensitivity improved from 0.86561 to 0.9554, and mean specificity improved from 0.82645 to 0.83116.

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 Rayvolve algorithm's performance specifically vary within the pediatric subgroup for different types of fractures (e.g., buckle, greenstick) compared to common adult fracture types?; What were the specific criteria and detailed methodology used for establishing ground truth (presence/absence and localization of fractures) in both the standalone performance datasets and the multi-reader multi-case (MRMC) reader study?; Given the expansion of the training dataset for the retrained algorithm from 115,000 to 150,000 radiographs to include more pediatric cases, what was the proportional representation of pediatric versus adult cases, and how did this influence model generalization and potential biases across different age groups?; The document mentions Rayvolve performs with high accuracy across different X-ray system manufacturers (e.g., Konica Minolta, IRay, AGFA, GEHC, Philips, Carestream); what quantitative measures or specific sub-analyses support this claim beyond the listed manufacturers, and are there known limitations or performance variations with specific imaging hardware, acquisition protocols, or detector types?; How does Rayvolve's deep learning model handle images with common artifacts, suboptimal quality (e.g., motion, under/over-exposure), or atypical patient positioning, and what is the documented impact of these factors on its diagnostic accuracy and the reliability of its fracture detection output in real-world clinical scenarios?