OC11.02: The importance of ultrasound preceding non‐invasive prenatal testing for fetal chromosomal abnormalities

Abstract

Objectives: The clinical workflow of the second trimester anomaly scan is not well studied and holds potential for efficiency improvement. We aimed to create a model for automatic anatomical description of full-length fetal anomaly scan videos using artificial intelligence. Methods: We prospectively recorded routine full-length second trimester anomaly scans, extracted short clips of important scan events by detecting video freeze, and image/clip save. For machine learning, we created a training dataset by manually labelling 12% of the scan events to one of 23 principal anatomical structures, trained a deep spatiotemporal neural network with the training dataset, cross-validated and applied the model to automatically label the rest of the scans. Finally, we retrospectively labelled a test dataset (48 scans) to compare with the automatically labelled scans. We report the model precision and workflow metrics. Results: 518 scans performed by 14 operators were analysed. The mean scan duration was 26.7 ± 15 minutes, and the mean number of scan events was 23.5 ± 14.4. The manual vs. automatic clips labelling agreement was 74.5%, ranging from 34% for placenta to 89% for heart. The brain, heart and spine were most often the first structure to be evaluated, in 18.8%, 17.6% and 17% of the scans, respectively. On average, 15% of the scan duration was dedicated to cardiac scanning, 10% to brain, and 7% to the spine (figure 1). Conclusions: Using big data, we present a model that describes how expert sonographers perform anomaly scans in a data science fashion. Understanding how operators scan and being able to measure the different operator elements will inform a better understanding of how to train operators, monitor learning progress, and enhance scanning workflow.