EP01.07: Robust and accurate measurement of semilunar valves in second trimester fetal ultrasound using an AI system

Measurements of the semilunar valves can assist in detecting congenital heart defects (CHD) in 2nd trimester (2T) fetal ultrasounds. However, such measurements are not universally performed during fetal ultrasound scans. We aimed at evaluating an AI system measuring aortic valve (AV) and pulmonary valve (PV) from fetal ultrasound video clips, both alone and after limited user review. We included 30 cases (18-24 weeks GA) including 10 CHD cases, recorded with a GE Expert 22 and Voluson E8. Each case consisted in one or two video clips featuring the left and right ventricular outflow tracts (LVOT, RVOT). The AI system measures AV and PV using entire video clips, and outputs a caliper of each valve on a most representative frame. The AI system was developed independently from this evaluation. Three experts each measured AV and PV independently. Experts were instructed to measure valve annulus internal size in long axis LVOT and transverse or high short axis RVOT views during systole, similar to the AI. Then, in order to evaluate users performing limited review of AI outputs, each expert performed “AI-aided” measurements of the same cases by reviewing calipers from the AI and adjusting them if needed, without changing frame. The mean intraclass correlation coefficient (ICC) was 0.678 between independent experts and AI, and 0.715 between independent experts and AI-aided measurements, higher than between pairs of independent experts (0.598). Therefore, AI and AI-aided measurements had a higher agreement with independent experts compared to independent experts with each other. The mean ICC was 0.920 between pairs of AI-aided measurements, showing that AI-aided measurements had reduced interobserver variability compared to independent experts. AI offers a promising method for accurate and robust measurement of semilunar valves in 2T fetal ultrasounds. Please note: The publisher is not responsible for the content or functionality of any supporting information supplied by the authors. Any queries (other than missing content) should be directed to the corresponding author for the article.