Automated interpretation of fetal cardiac function evaluation from the echocardiogram

Purpose: Prenatal assessment of fetal cardiac function is crucial for predicting neonatal outcomes, yet manual measurements from fetal echocardiograms are labor-intensive and prone to human error. To address these challenges, we developed a fully automated artificial intelligence (AI) workflow to estimate fetal cardiac function parameters from echocardiograms. Methods: The AI workflow implement a deep learning architecture with capability of real-time detecting and segmenting the potential cardiac structures from the whole echocardiogram. The AI workflow considers a quality control by the predictive confidence and dynamic trends of detected structures, and then effectively provide multiple measurement parameters from the segmentation. We developed and validated the AI workflow using a internal dataset of 52,942 annotated images from 1,940 echocardiograms with the normal singleton fetal hearts at the First Affiliated Hospital of Sun Yat-sen University. We also validated the segmentation performance of the workflow in two external normal dataset (129 echocardiograms at the Seventh Affiliated Hospital of Sun Yat-sen University and 116 echocardiograms at Zhongshan City People’s Hospital), and one internal abnormal dataset at the First Affiliated Hospital of Sun Yat-sen University. We validated the automatic measurements of workflow against manual measurements in the ultrasound instrument with or without the fetal Heart Quantification software measurements by two expert sonographers. The AI workflow further establishes a Z-score model with dynamic consideration of the gestational age information and fetal biometric parameters to standardize the evaluation system of fetal cardiac function. Results: The segmentation of the workflow was accurate, with a mean Dice similarity coefficient greater than 92% and a mean intersection-over-union greater than 85%, over both the internal and external test dataset. Automated measurements showed strong agreement with manual and Fetal HQ measurements, with intra-class correlation coefficients ranging from 0.817 to 0.995 and mean absolute errors ranging from 0.009 to 5.415. Bland-Altman limits of agreement indicated good agreement between automated and human measurements. The mean individual equivalence coefficients for all parameters were less than 0, indicating lower variability in automated measurements compared to manual and fetal HQ measurements. Conclusion: The AI-based workflow can accurately segment fetal cardiac structures and systematically quantify cardiac function parameters with high precision and reliability, comparable to expert manual measurements. This automated approach provide accurate, simple, efficient and repeatable intelligent tool for cardiac function quantification in clinical practice.