FETOLY-HEART

FETOLY-HEART utilizes a supervised deep learning algorithm to analyze fetal ultrasound images in real-time, automatically detecting five standard cardiac views and 52 associated quality criteria. The system serves as a concurrent reading aid for healthcare professionals, aiming to improve the completeness and standardization of fetal heart examinations during the 2nd and 3rd trimesters.

FETOLY-HEART is a software designed to assist healthcare professionals in performing real-time fetal heart ultrasound examinations during the second and third trimesters of pregnancy (17 to 40 weeks gestational age). Its primary purpose is to analyze fetal ultrasound image sequences using machine learning techniques, specifically a frozen deep learning algorithm trained via supervised learning, to automatically detect five standard fetal heart views (Abdomen, Four chamber, Left Ventricular Outflow Tract, Right Ventricular Outflow Tract, Three vessels view) and 52 predefined anatomical quality criteria within those views. The system, installed on hardware connected to an ultrasound machine via HDMI, processes images in real-time and aids in assessing the completeness of the examination against established guidelines by displaying which heart views and quality criteria have been verified. The software provides two main functionalities: a real-time "completeness overview" indicating verified views and criteria, and a "completeness illustration" which selects an image subset as visual evidence for clinician review, with optional display of quality criteria localization. Performance testing on a dataset of 2,288 images from 480 patient cases demonstrated high efficacy: fetal heart view detection achieved sensitivities and specificities generally above 97% (e.g., Abdomen view sensitivity 0.976, specificity 0.998), exceeding the ≥85% acceptance criteria. Detection of the 52 quality criteria met sensitivity and specificity targets of ≥90%, with most criteria achieving values above this threshold. Localization of these criteria via bounding boxes also met the mean Intersection over Union (IoU) target of ≥50%, with many achieving higher values.

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 "frozen" nature of the deep learning algorithm affect its adaptability to new ultrasound machine image characteristics or evolving imaging protocols not covered in the initial training dataset?; What are the specific computational requirements (e.g., processing power, memory) for the hardware device running FETOLY-HEART to ensure real-time analysis without latency during an examination?; The performance study mentions subgroup analysis; were there any statistically significant differences in performance for specific fetal abnormalities, and how might this impact its utility in high-risk pregnancies?; Could the system's "completeness illustration" feature be expanded to rank or score the quality of the selected images, beyond just verifying presence, to further guide less experienced sonographers?; How robust is the algorithm in differentiating between closely related anatomical structures or views when image quality is suboptimal due to factors like maternal BMI or fetal positioning?