ADAS 3D

ADAS 3D is a medical image processing software that employs advanced algorithms, incorporating machine learning, to post-process cardiac MR and CTA images for detailed 3D visualization and quantification of cardiac structures and myocardial scar. The system aims to enhance clinical decision-making in cardiovascular disease diagnosis and support the planning of electrophysiology procedures by offering semi-automated analysis of cardiac anatomy and pathology.

ADAS 3D is a stand-alone software tool designed for the post-processing of cardiovascular Magnetic Resonance (MR) and Computed Tomography Angiography (CTA) images, provided in DICOM format. Its core function is to aid qualified medical professionals in the non-invasive calculation, quantification, and three-dimensional visualization of cardiac imaging data, thereby supporting diagnostic decision-making for patients with cardiovascular disease, especially those with myocardial scar from ischemic or non-ischemic heart conditions. The software enables detailed analysis of myocardial fibrosis enhancement, including its 3D distribution, quantification of total enhancement volume within the Left Ventricle (LV) and Left Atrium (LA), and the characterization of intermediate signal intensity corridors in the LV. Additionally, it supports quantification of LV and LA wall thickness from CTA images, visualization of other 3D anatomical structures from CTA and Magnetic Resonance Angiography (MRA), and is intended for use in pre-planning and during electrophysiology procedures, exporting data compatible with catheter navigation systems. A significant component of ADAS 3D is its utilization of machine learning-based features for the semi-automatic initialization and segmentation of cardiac structures, such as the LV, LA, aorta, and coronary arteries from both CTA and MR images; these functionalities were either newly introduced or improved in this iteration of the device. Non-clinical performance testing of these algorithms involved training on diverse, anonymized DICOM datasets sourced from European hospitals, followed by testing on distinct datasets from both US and OUS hospitals, encompassing a variety of CT/MRI scanners. While the majority of performance metrics (e.g., Dice Metric, Mean Surface Distance, Hausdorff Distance) successfully met predefined non-inferiority acceptance criteria, four specific sub-tests related to mean surface or perpendicular distances for LA endocardium (from CTA), LV endocardium and epicardium (from 2D LGE-MRI), and LA (from 3D LGE-MRI) did not meet their stringent numerical thresholds. However, these discrepancies were characterized by the manufacturer as sub-pixel and thus considered acceptable, with the overall performance of the machine learning features deemed to be in line with 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 performance of the machine learning segmentation for LA wall thickness from CTA, which was changed from manual to semi-automatic, compare quantitatively to the previous manual method or an expert-defined ground truth?; Given the four performance metrics that did not meet strict non-inferiority criteria (LA ENDO MSD from CTA, LV ENDO APD and LV EPI APD from 2D LGE-MRI, and LA MSD from 3D LGE-MRI), what is the potential clinical impact, if any, of these reported sub-pixel discrepancies on downstream tasks like electrophysiology procedure planning or scar quantification accuracy?; The training data for machine learning features was anonymized, preventing inclusion of patient demographics like gender, age, or ethnicity. What steps were taken, or are planned, to evaluate and ensure the algorithms generalize robustly across diverse demographic populations potentially underrepresented in the European hospital training datasets?; The document states, 'The clinical significance of using ADAS 3D to identify arrhythmia substrates for the treatment of cardiac arrhythmias (e.g., ventricular tachycardia) or risk stratification has not been established.' What specific clinical studies are planned or currently underway to investigate and establish this clinical utility?; How does ADAS 3D's performance, particularly its machine learning features, vary with different levels of input MR/CT image quality and resolution, and are there defined minimum image acquisition parameters recommended for optimal results?