OC10.05: An artificial intelligence system that can correctly identify fetal ultrasound imaging planes throughout gestational age

Elizabeth Bradburn, Yazeed Mohammad, J. Alison Noble, Aris T. PapageorghiouUltrasound 2020 · DOI

Abstract

To assess the accuracy of an artificial intelligence (AI) algorithm at identifying fetal imaging planes on prenatal ultrasound, throughout gestational age. Anonymised fetal ultrasound images, acquired at 15 to 40 weeks of gestation, from a large image database were analysed in this study. This included images of head circumference (HC), abdominal circumference (AC), femur length (FL), amniotic fluid index (AFI) and miscellaneous images. An AI algorithm (ScanNav, trained primarily on routine anatomy scans from 18-24 weeks), was applied to each image and categorised it into one of five categories: HC, AC, FL, AFI and others. A random ten percent per view of all images were manually reviewed by one experienced clinician to determine the accuracy of this classification. The performance of the AI algorithm, according to the gestational age at the time of the scan, was assessed. Over 450,000 images were assessed by ScanNav, and a random 10% (45,036 images) underwent clinical review. The accuracy of the AI algorithm in categorising the images into standard biometry planes (HC, AC, FL and AFI) was 99.1%. This was 99.6%, 99.1%, 97.9% and 99.9% for HC, AC, FL and AFI, respectively. Across all categories the algorithm performed best between 20+0 and 35+6 weeks, with an accuracy that remained stable throughout that gestational age range (97%). There was a slight drop in accuracy below 20 weeks and above 36 weeks, to 94% and 96% respectively. Although trained primarily on routine anatomy scans from 18-24 weeks, the ScanNav AI algorithm is 99.1% accurate in image categorisation of basic biometry planes. Accuracy was maintained over a very broad range of gestational ages. This may reflect that clinically these standard planes remain consistent in appearance throughout gestation. Improvement at the extremes of GA may be possible by further algorithm training.
BibTeX
@inproceedings{bradburn2020oc,
  author = {Elizabeth Bradburn and Yazeed Mohammad and J. Alison Noble and Aris T. Papageorghiou},
  title = {{OC10.05: An artificial intelligence system that can correctly identify fetal ultrasound imaging planes throughout gestational age}},
  booktitle = {Ultrasound},
  year = {2020},
  doi = {10.1002/uog.22269},
}

Loading related content...

Figure