Sonio Detect

Sonio Detect is a machine learning-powered Software as a Service designed to enhance fetal ultrasound examinations by automatically detecting views, identifying anatomical structures, and verifying image quality criteria and characteristics in real-time. This system acts as a concurrent reading aid for healthcare professionals, utilizing AI, biometrics, and colorimetry to improve the consistency and completeness of obstetric imaging across trimesters.

Sonio Detect is a Software as a Service (SaaS) solution designed to assist healthcare professionals, including sonographers, OB/GYN Maternal-Fetal Medicine specialists, and fetal surgeons, during routine fetal ultrasound examinations across all three trimesters (gestational age 11-37 weeks). Its primary purpose is to serve as a real-time concurrent reading aid, helping ensure the completeness of examinations and adherence to imaging protocols. The system operates by receiving DICOM fetal ultrasound images and clips from an ultrasound machine via a locally installed Edge Software, which uploads them to Sonio's cloud. Sonio Detect then employs machine learning algorithms to automatically detect standard fetal views, identify anatomical structures within these views, and verify predefined quality criteria and characteristics, such as placenta location or fetal sex. The underlying science of Sonio Detect relies on artificial intelligence, incorporating methodologies like biometrics analysis and colorimetry for 3D and Doppler images. The software is designed to recognize numerous specific views (e.g., transthalamic, 4 chambers, abdominal circumference), anatomical structures (e.g., cavum septum pellucidum, ventricles, kidneys), and quality criteria (e.g., presence of specific structures, absence of others) across different trimesters, as well as characteristics like fetal sex and placental location. Performance data from bench testing on a dataset of 36,769 fetal ultrasound images demonstrated high sensitivity and specificity for these automated detection and verification tasks. For instance, automatic detection of T2/T3 fetal ultrasound views (18 types) showed a sensitivity of 0.937, and detection of fetal brain anatomical structures (8 types) at T2/T3 showed a sensitivity of 0.934 and specificity of 0.949. The system provides visual cues for unverified items and allows users to override its automated findings.

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 Sonio Detect's performance in automatically verifying "characteristics" (e.g., fetal sex, placenta location) compare to its performance in verifying "quality criteria" related to anatomical presence/absence, and what are the clinical implications of potential discrepancies?; What are the specific machine learning models or architectural components used for view detection versus anatomical structure identification, and how are they trained to handle variations in image quality or fetal presentation across different trimesters and ultrasound manufacturers (GE, Canon, Philips, Samsung)?; The system allows user overrides; what mechanisms are in place to learn from these overrides to potentially improve the algorithm's performance over time, and how is the integrity of the "standardized quality criteria" maintained if frequent overrides occur for specific criteria?; Considering the subgroup analysis (BMI, maternal age, gestational age, race/ethnicity), were any specific subgroups identified where the algorithm's sensitivity or specificity was notably different, and what steps are being taken to address or understand these variations?; How does the real-time "concurrent reading aid" functionality integrate into typical clinical workflows, and what is the measured impact on examination time or the rate of complete and protocol-adherent studies compared to unassisted examinations?