Automatic Anatomy Recognition (AAR)

The Automatic Anatomy Recognition (AAR) system is a cloud-based, software-only medical device employing deep learning to automatically generate anatomical contours of organs at risk from non-contrast CT studies for radiation therapy planning in adult head & neck and thoracic oncology cases. It operates without human intervention for initial segmentation and requires DICOM 3.0 compatible third-party systems for contour display and modification, with its accuracy validated by DICE similarity coefficients and Hausdorff Distance metrics.

The Automatic Anatomy Recognition (AAR) system by Quantitative Radiology Solutions, LLC, is a software-only medical device operating on a Linux, cloud-based platform, intended for technicians and trained physicians. Its core purpose is to automatically derive contours of anatomical structures, specifically "organs at risk" (OAR), from non-contrast computed tomography (CT) studies. These DICOM 3.0 compliant contours serve as input for radiation treatment planning systems. AAR is designed for adult patients (18 years and older) and focuses on anatomical structures within the head & neck and thoracic body regions. It functions independently of the specific radiation treatment delivery method (e.g., photons, protons), modality (e.g., 3D-CRT, IMRT), therapeutic intent (e.g., curative, palliative), or the underlying disease process being treated. AAR employs deep learning algorithms to automatically process CT studies and generate contours without human intervention during the segmentation phase; it can receive, add/edit/delete, transmit, and input/export medical images and DICOM data. A crucial operational aspect is that AAR does not include tools for modifying the generated contours. Any required adjustments or display of contours must be performed using compatible third-party treatment planning systems, which must also adhere to DICOM 3.0 standards, similar to the CT scanners. The system's performance in automated segmentation accuracy was validated through tests using DICE similarity coefficients and mean 95% Hausdorff Distance (HD) calculations. Furthermore, comprehensive software verification and validation were conducted following the IEC 62304 standard, covering unit, integration, and system-level testing.

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. What specific deep learning architecture is employed by AAR, and how was its training dataset curated to ensure robustness across diverse patient populations and CT scanner variations within the head & neck and thoracic regions?; Could you provide the quantitative results (e.g., mean DICE scores, Hausdorff distances with standard deviations) for specific organs at risk (e.g., spinal cord, lungs, parotid glands) achieved during the segmentation performance tests?; Given that AAR does not allow for contour modification, what is the anticipated clinical workflow adjustment, and what is the typical percentage of contours requiring modification in a downstream system based on pre-market evaluations?; How does the cloud-based deployment of AAR address data security, patient privacy (HIPAA compliance in the US context), and system latency, particularly when processing large CT datasets?; What are the documented failure modes or scenarios where AAR's automated contouring accuracy might be compromised (e.g., presence of metallic artifacts, significant anatomical anomalies, poor image quality), and are there any built-in quality indicators for the generated contours?