Automated and accurate view classification of fetal echocardiography using convolutional neural networks

Accurate view classification is imperative for the prenatal precision diagnosis of congenital heart disease (CHD) but challenges remain, including suboptimal imaging quality due to noise and artifacts, along with a high reliance on extensive operator or diagnostician expertise, which complicates this process. We aimed to analyze several convolutional neural network (CNN)-based frameworks for the automated classification of eight standard fetal echocardiographic views.

The study utilized fetal echocardiographic data collected from 953 second- and third-trimester fetuses between July 2018 and October 2019. Among them, 10,032 images were randomly selected for training and validation, while 2,071 images were used to assess the framework's accuracy. Various advanced CNNs were trained to identify the recommended set of eight fetal echocardiographic views. Model performance was compared with the ground truth, followed by an in-depth analysis of results obtained from the optimal model. Gradient-weighted class activation mapping (Grad-CAM) was employed to visualize model decision-making.

The ConvNeXt-based model achieved the highest performance with an overall test accuracy of 97.78% and an F1-score of 96.08% in classifying the eight preselected views. The accuracy of most views exceeded 95%. Although image quality impacted the model's classification performance (P<0.001), this effect remained within acceptable limits for clinical applicability. Data visualization further elucidated that the model identified similarities across different views and utilized clinically relevant features for classification.

CNNs have emerged as a promising tool for facilitating accurate and efficient view classification of fetal echocardiograms. This study paves the way for the future utilization of artificial intelligence (AI) to automate fetal heart screening and potentially support the diagnosis of CHD.