AI assistance for fetal ultrasound interpretation in a multi-reader study

Abstract Prenatal detection rates of congenital anomalies range from 25 to 90%. Artificial intelligence (AI) may improve detection by supporting physicians during image interpretation. We performed a multireader, multicase retrospective study of Sonio Suspect (FDA‑cleared, K243614; Paris, France) detecting 8 fetal abnormal ultrasound findings on still images: malposition of the great vessels, absence or unusual size of at least one of the three vessels, disequilibrium or absence of at least one of the two ventricles, thoracic situs inversus, abdominal situs inversus, non-visibility of a single stomach bubble or abnormally big stomach, absence of the cavum septum pellucidum, absence of the corpus callosum. Thirteen U.S. physicians (maternal-fetal medicine, obstetrics/gynecology, radiology) reviewed 750 images (250 with abnormal findings) once unassisted and once with AI. With AI assistance, AUC presented a mean improvement of 21.9 percentage points (68.9% to 90.9%, p < 0.001). With AI, sensitivity was overall improved (88.5% vs. 54.2%, p < 0.01), as was specificity (87.9% vs. 80.7%, p < 0.01), inter-reader agreement (72% vs 26%) and interpretation time (23 vs 40 s per image). AI assistance significantly improved reader performance in detecting fetal abnormalities without compromising specificity.