A deep neural network could accurately detect AVSD in 2nd trimester fetal heart ultrasound video clips.

AVSDs are often undetected before birth, with an impact on morbidity and mortality. In addition, detecting AVSD may allow diagnosing genetic disorders (such as Down syndrome), which are frequently associated.

Here, we aim at evaluating whether a deep neural network (DNN) could identify AVSD in 2nd trimester (2T) fetal ultrasound video clips.

Patients with single pregnancy who had an echocardiography performed at one center (18–25weeks GA) were included retrospectively starting from Jan 1, 2021. Based on clinical records, we included consecutive cases of partial or complete AVSD, and consecutive negative cases referred due to family history. This inclusion criterion was used for negative cases to be more representative of the general population, in a center with a high prevalence of CHD since it only receives patients referred for echocardiography.

Cases were reviewed by one of two fetal echocardiography experts to confirm that the presence or absence of AVSD was documented in at least one video clip. Patients with no such video clip were excluded.

The DNN takes as input all the recorded video clips of a given examination and outputs the absence or presence of AVSD, or an “inconclusive” output if its confidence is low. The DNN was trained to detect AVSD, as seen on the four-chamber view, on patients not included in the evaluation.

We included 26 cases with AVSD and 129 cases without. The DNN achieved an AUC of 97.1%, a sensitivity of 86.4% (95% CI: 66.7–95.3) and a specificity of 95.2% (95% CI: 90.0–97.8), after excluding inconclusive diagnosis. The DNN predicted a conclusive diagnosis in 95.5% of cases (Fig. 1).

A DNN could accurately identify AVSD in 2T fetal echocardiography. These results establish the groundwork for efficient and accurate AI-assisted fetal ultrasound heart screening.