Abstract
Prenatal assessment of fetal cardiac function is crucial for predicting neonatal outcomes, yet manual echocardiographic measurements are labor-intensive and subjective. We developed a fully automated artificial intelligence (AI) workflow for estimating fetal cardiac function parameters from echocardiograms. The workflow integrates a deep learning model for real-time detection and segmentation of cardiac structures, followed by quality control and geometric calculation. It was developed and validated using an internal dataset of 52,942 annotated images from 1940 normal fetal echocardiograms, with further testing on two external normal datasets (245 echocardiograms) and one internal abnormal dataset (83 echocardiograms). Performance was compared against manual and Fetal Heart Quantification (Fetal HQ) measurements, and a dynamic Z-score model referencing gestational age and fetal biometrics was established. The AI achieved accurate segmentation, with mean Dice similarity coefficients >92% and mean intersection-over-union >85% across all test datasets. It exhibited higher intraclass correction coefficients and R-values relative to experts than inter-observer variability, alongside smaller mean absolute error and limits of agreement. The mean individual equivalence coefficients of all cardiac function parameters were below zero, indicating lower variability than manual or Fetal HQ. These results demonstrate that our fully automated AI workflow enables accurate, efficient, and reproducible quantification of fetal cardiac function, supporting its potential for standardized clinical application.