Application of a multi-task AI model for standard plane recognition and structure segmentation in fetal cardiac ultrasound

LI Huilian; LI Furong; LIU Peizhong; HE Shaozheng

doi:10.12122/j.issn.1674-4500.2025.12.04

Abstract：Objective To integrate artificial intelligence with prenatal ultrasound imaging by developing a deep-learning multi-task model that simultaneously identifies standard fetal cardiac planes and performs precise instance segmentation of key anatomical structures.Methods A total of 3312 fetal cardiac ultrasound images were collected from 1300 singleton pregnancies at 18-24 weeks of gestation in the Second Affiliated Hospital of Fujian Medical University from January 2021 to July 2023,and all images were jointly annotated by three associate chief sonographers.The dataset covers five standard cardiac planes,apical four-chamber(4CH),three-vessel(3VV),three-vessel-and-trachea(3VT),right ventricular outflow tract(RVOT)and left ventricular outflow tract(LVOT),together with ten critical structures,including the left/right ventricles,left/right atria,pulmonary artery,main pulmonary artery,ascending aorta,aorta,trachea and superior vena cava.YOLOv11 was adopted as the backbone network for instance segmentation.An additional classification branch was attached to the detection head to realize joint learning of plane recognition and structure segmentation.Results The proposed model achieved an AUC of 0.976 for plane recognition and an mAP of 0.937 for instance segmentation.Compared with single-task models performing only classification or segmentation,the multi-task framework demonstrated superior overall performance.Conclusion The YOLOv11-based multi-task learning model accurately recognizes standard fetal cardiac planes and delineates key anatomical structures,offering considerable potential to enhance the efficiency and accuracy of prenatal screening for congenital heart anomalies.