Binary Classification of Fetal Congenital Heart Diseases Using Deep Learning and Temporal Feature Integration in Echocardiographic Videos

Vasquez-Solis L.; Anco-Copaja E.; Aures-Garcia A.; Huertas-Tacchino E.; Balvin-Landeo A. · Proceedings of the 2025 IEEE 5th International Conference on Advanced Learning Technologies on Education and Research, ICALTER 2025 (2025) · DOI: 10.1109/icalter69698.2025.11355125

Resumen

Congenital heart diseases (CHDs) represent one of the most prevalent birth abnormalities, occurring in approximately 8-12 per 1,000 live births globally. Early prenatal detection through fetal echocardiography is critical for improving perinatal outcomes. This research presents a deep learning methodology for automated binary classification of fetal CHDs using a hybrid architecture combining ResNet50 convolutional neural networks with bidirectional long shortterm memory (Bi-LSTM) networks to leverage both spatial and temporal features from echocardiographic videos. The methodology employs transfer learning from ImageNet pretrained weights and implements robust 5-fold stratified crossvalidation. The dataset comprises 132 echocardiographic videos from the 4-chamber view (72 healthy, 60 CHD) acquired from multiple ultrasound vendors. The model achieved AUC-ROC of 0.941 pm 0.046, sensitivity of 95.0 % pm 6.7 %, specificity of 83.4 % pm 6.8 %, and accuracy of 88.6 % pm 4.8 %, demonstrating clinically relevant performance for binary CHD screening.

Instituciones: Universidad Nacional Mayor de San Marcos; Universidad Peruana de Ciencias Aplicadas