Sonix Health

Sonix Health is a software tool designed to enhance echocardiographic analysis by automating the quantification and reporting of cardiac parameters from ultrasound images. It employs a two-step deep learning algorithm to identify imaging views and perform automated measurements on DICOM-compliant B-mode, M-mode, and Doppler data, with functionalities for manual verification and adjustment by physicians.

Sonix Health is a software-only medical image management and processing system designed for quantifying and reporting echocardiography in adult populations, intended for use by or on the order of a licensed physician. It processes DICOM-compliant medical images acquired from ultrasound devices using B (2D), M, Pulsed-wave Doppler, and Continuous-wave Doppler modes. The system automates previously manual quantification tasks by checking ultrasound multiframe DICOMs, performing classification and automatic measurement, allowing users to verify and manually adjust results, and providing an analysis report, thereby aiming to increase user efficiency. The core of Sonix Health is a two-step algorithm: an initial identification model determines the echocardiographic view, followed by a view-specific deep learning algorithm (categorized into B-mode, M-mode, and Doppler algorithms) for anatomical segmentation and measurement. This software is installed on customer PC hardware and operates within a local network. Performance testing, using 2,744 images from diverse sources and equipment, demonstrated an average accuracy of 98.22% for view recognition and a 93.98% average correlation coefficient for measurements compared to manual expert assessments.

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 two-step deep learning algorithm in Sonix Health adapt to variations in image acquisition protocols or patient-specific anatomical anomalies not extensively covered in its training dataset?; What is the performance degradation, if any, of the view classification and automated measurement features when processing images from ultrasound systems with significantly lower resolution or higher noise levels than those used in the validation studies?; Can the system provide confidence scores for its automated measurements and view classifications, and how are these scores correlated with the likelihood of requiring manual correction by a clinician?; Beyond the listed semi-automated measurements, what is the roadmap for expanding the range of quantifiable cardiac parameters, particularly for complex or less common pathologies?; How does the system's performance in identifying and measuring specific cardiac structures (e.g., left ventricular dimensions, Doppler velocities) compare across different echocardiographic views, and are there particular views where its accuracy is more variable?