Identification of abnormal ventricular asymmetry in 2nd trimester fetal ultrasound using artificial intelligence

Abstract

Abstract Introduction Prenatal ultrasound screening for congenital heart disease (CHD) suffers from poor detection yield (10%-56%)[1, 2], with a strong impact on neonatal morbidity and mortality. The need for expertise in fetal cardiology and the high volume of screening cases limit the current detection rates. Computer-assisted prenatal screening could support clinicians in improving diagnostic yield, including for abnormal ventricular asymmetry, which is often hard to discriminate from physiological asymmetry. Here we aim at evaluating a deep neural network (DNN) for detection of abnormal ventricular asymmetry, which can be linked to multiple types of CHD, including coarctation of the aorta, pulmonary or aortic valve stenosis, ventricular hypoplasia or univentricular heart. Purpose To determine if a novel DNN can identify cases of abnormal ventricular asymmetry in 2nd trimester fetal ultrasound videos. Methods We used as an endpoint the presence of any abnormality causing a significantly abnormal asymmetry of the ventricles as seen in the fetal ultrasound. Consecutive patients who had an echocardiography performed at one center were included retrospectively starting from 2022, January 1st. Inclusion criteria were single pregnancy between 18 and 25 weeks of gestation. To have a high number of positive patients, patients were included separately if they were expected to be positive (n=35) or negative (n=75) to the endpoint, based on their medical record. Each case was reviewed by one of two physicians expert in fetal echocardiography interpretation to confirm the presence or absence of the endpoint. For each patient, a single video showing an interpretable four chamber view was selected. Patients with no such video were excluded. The DNN takes as input fetal ultrasound videos, and predicts the absence or presence of the endpoint for each video. If its confidence is low, the DNN predicts an "inconclusive" output instead. The DNN has a ResNet-50 architecture[3] adapted to videos and was trained on data from patients not included in the evaluation. The detection performance of the DNN and the proportion of conclusive outputs were assessed. Results Overall, 37 positive and 66 negative ultrasound videos were included. The DNN achieved an AUC of 75.5%, a sensitivity of 71.4% (54.9-83.7, 95%CI) and a specificity of 64.4% (51.7-75.4, 95%CI) in identifying cases of abnormal ventricular asymmetry, after excluding inconclusive diagnosis. The DNN predicted a conclusive diagnosis in 91.3% of cases. Conclusion A DNN could accurately identify cases of abnormal ventricular asymmetry in fetal heart ultrasound videos. Importantly, this evaluation was performed in a population of patients who had an echocardiogram, and which therefore contains more abnormal cases compared to patients undergoing standard screening. This work lays the ground for diagnostic support to enable a more efficient, accurate, and scalable analysis of fetal heart ultrasounds.