AI-Rad Companion (Musculoskeletal)

AI-Rad Companion (Musculoskeletal) VA20 is an advanced image processing software that utilizes deep learning algorithms to automatically segment, label, and measure vertebrae from adult CT scans, providing quantitative data like vertebral heights and Hounsfield Unit values. Its purpose is to enhance clinicians' ability to evaluate musculoskeletal diseases of the spine by offering improved segmentation accuracy, flexible deployment options (cloud and on-premise), and standardized DICOM reporting.

The AI-Rad Companion (Musculoskeletal) SW version VA20 is an image processing software designed to assist clinicians in evaluating and assessing musculoskeletal diseases, specifically focusing on the spine. It analyzes previously acquired Computed Tomography (CT) DICOM images from adult patients, employing deep learning algorithms for its core functionalities. These functionalities include the segmentation and labeling of vertebrae, measuring vertebral heights (flagging critical differences), and determining the mean Hounsfield Unit (HU) values within vertebral volumes of interest. This version represents an enhancement to a prior iteration, featuring an improved AI algorithm retrained for greater vertebrae segmentation accuracy. The scientific basis of the device lies in its deep learning image-to-image network for 3D segmentation. Key enhancements in version VA20 include improved DICOM reporting capabilities, supporting both human and machine-readable formats (including TID 1500), and expanded deployment options to include on-premise "edge" processing alongside existing cloud deployment, while maintaining cloud connectivity for monitoring. Non-clinical performance validation on 140 CT datasets (1,553 thoracic vertebrae) demonstrated that the software met its acceptance criteria: a mislabeling/missing measurement failure rate of 8.6% (below the 10% threshold), and inter-reader variability for height measurements within limits of agreement at 95.5% for slices ≤1.0mm and 92.6% for slices >1.0mm. Consistency across critical sub-groups was also maintained or improved compared to the predicate device.

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 enhanced AI algorithm's performance in vertebrae segmentation compare quantitatively (e.g., Dice score, Jaccard index) to the previous version (K193267) beyond general statements of "improved accuracy"?; What specific types of musculoskeletal diseases or conditions benefit most from the quantitative and qualitative analyses provided by this software, and are there any known limitations for certain pathologies?; What was the range and distribution of Hounsfield Unit values in the validation dataset, and how does the software's accuracy in HU measurement vary across different bone densities (e.g., osteoporotic vs. normal bone)?; Could the on-premise "edge" deployment option introduce variability in processing times or results compared to the centralized cloud deployment, and what are the minimum hardware specifications for the edge system?; What were the "critical differences" in vertebral heights that the software is designed to indicate, and how were these thresholds established and validated for clinical significance?