OC10.11: The data science of obstetric ultrasound: automatic analysis of full‐length anomaly scans using machine learning algorithms

Abstract

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. 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. 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). 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. Supporting Information 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.

Ähnliche Arbeiten