Diagnosis of congenital heart diseases in children from 2D and 3D sonography using convolutional neural networks: A scoping literature review

Abstract

Congenital heart diseases (CHDs) are the commonest congenital anomalies and a leading cause of neonatal morbidity and mortality worldwide. They affect approximately 1 % of live births globally. Prenatal detection with fetal echocardiography allows for timely referral and intervention. However, screening performance remains uneven because of operator dependence and resource limits. Convolutional neural networks (CNNs) offer automated image interpretation and have been widely applied to 2D and 3D fetal ultrasound (US) in recent years.

This scoping review maps CNN applications for prenatal CHD across 2010–September 2025. In adherence to the PRISMA-ScR methodology, we screened 845 records from PubMed, IEEE Xplore, Google Scholar, ScienceDirect and preprint servers and included 29 studies for synthesis. The volume of articles published on the topic rose sharply after 2020, with most datasets drawn from Asia and North America and no representation from sub-Saharan Africa. The most commonly reported architectures were DenseNet variants, U-Net families for segmentation, YOLO (You Only Look Once) variants for real-time detection, and ensemble hybrids for classification.

Reported internal performance is high in many studies, with top models achieving near-expert discrimination with AUCs (Area Under the Curve) up to 0.99. In our cohort, the mean AUC across studies reporting that metric was 0.911 (Standard Deviation-SD = 0.09; 95 % CI: 0.86–0.96; n = 11). Average sensitivity and specificity across studies reporting these metrics were 0.92 (SD = 0.04; n = 13) and 0.91 (SD = 0.04; n = 7), respectively. However, external and multi-centre validation remains limited, and performance commonly falls when models are tested on unseen centers or different scanner types. Methodological gaps include small or imbalanced lesion classes, inconsistent patient-level splitting, sparse reporting of deployment constraints, and limited fairness analyses.

Recent work shows promising directions such as multimodal fusion of B-mode and Doppler, lightweight networks, on-device optimization, federated and privacy-preserving training, and prospective protocols for multi-centre evaluation. To translate CNNs into clinical screening, we recommend coordinated efforts to build diverse, well-annotated repositories, adopt strict patient-level validation and external testing, report deployment metrics, and run prospective, pragmatic trials that measure clinical outcomes and health-economic impact. These steps can move CNNs from high-performance demonstrations to dependable tools that improve prenatal CHD detection equitably.