OP02.04: Automated segmentation of fetal heart three‐vessel view ultrasound video clips to facilitate prenatal assessment of congenital heart defects

To present an artificial intelligence (AI) method that characterises fetal heart anatomical structures and segments the fetal heart anatomy in the three-vessel view (3VV) on ultrasound videos. Sonographers captured a video clip by performing a transverse sweep through the fetal heart including images of situs, four-chamber view, left and right ventricular outflow tract, 3VV, and three-vessel tracheal view. Our method automatically segments the 3VV and characterises its structures by analysing morphological features. We manually segmented the pulmonary artery (PA), aorta (Ao), and superior vena cava (SVC) in 1,734 frames (35 videos) to support the understanding of morphology when performing the sweep. The method combines three networks with a refinement algorithm. An initial network defines the first frame along with the segmentation of the vessels. A second network characterises the vessels' morphological changes over the sequence of frames in the video. This information is fed into the third network to generate segmentations. The refinement algorithm enhances segmentations by filling in gaps where necessary. We report results in 664 testing frames (15 videos) using the intersection over union metric. Our method achieves an average of 89.5% (99.5% for PA, 85.0% for Ao, and 84.1% for SVC) and performs at 31 frames per second, which is superior to the state-of-the-art method by 7.1% (2.0% for PA, 9.4% for Ao, and 10.0% for SVC). Ultrasound-based assistive tools aim to help less experienced sonographers in scan interpretation by offering automated analysis. Our solution performs accurately at a high frame rate, making it suitable for real-time implementation. This could be used in detecting 3VV conotruncal abnormalities in ultrasound videos of the fetal heart.