OC10.03: Computer‐assisted low‐cost point of care ultrasound: an intelligent image analysis algorithm for diagnosis of malpresentation

Alice Self, Q. Chen, J. Alison Noble, Aris T. PapageorghiouUltrasound 2020 · DOI

Abstract

Undiagnosed breech presentation carries an increased risk of maternal and neonatal morbidity and mortality. Point of care ultrasound (POCUS) linked with automation technology could enable community midwives to accurately diagnose presentation. The aim was to establish if fetal malpresentation can be correctly identified in single sweep ultrasound videos. Women recruited to the computer-assisted low-cost point of care ultrasound (CALOPUS) study had a single sweep ultrasound scan from the pubic symphysis to the umbilicus during anomaly or growth scans. Fetal anatomical structures were manually annotated in videos from 172 subjects by fetal ultrasound experts and categorised as cephalic (C) or non-cephalic (NC). Algorithms were developed and trained to classify C versus NC presentation on 52 annotated videos. The algorithms' accuracies in presentation classification were tested on the 120 videos not used in training. The best performing algorithm integrated a deep convolutional neural network, responsible for detecting fetal anatomical structures, with a clinical decision-making matrix, based on the timing order when fetal structures were detected in the video. The overall accuracy was 94.2%; and rose to 100% when gestational age was >30 weeks. Cohen's kappa was 0.88 (95% CI 0.80-0.97). OC10.03: Table 1. 2x2 table of prediction accuracy 51 (92.7%) 19-22 weeks: 50 27-29 weeks: 0 31-34 weeks: 0 35-37 weeks: 1 4 (7.3%) 19-22 weeks: 4 27-29 weeks: 0 31-34 weeks: 0 35-37 weeks: 0 3 (4.6%) 19-22 weeks: 2 27-29 weeks: 1 31-34 weeks: 0 35-37 weeks: 0 62 (95.4%) 19-22 weeks: 44 27-29 weeks: 2 31-34 weeks: 2 35-37 weeks: 14 The machine learning algorithm detected fetal malpresentation with a high degree of accuracy. Use of a POCUS device to detect malpresentation in term fetuses could be a low cost way to initiate referral for appropriate management. Although the proposed algorithm showed 100% accuracy at gestational ages >30 weeks, it needs further verification on a larger dataset of fetuses at 35-37 weeks.
BibTeX
@inproceedings{self2020oc,
  author = {Alice Self and Q. Chen and J. Alison Noble and Aris T. Papageorghiou},
  title = {{OC10.03: Computer‐assisted low‐cost point of care ultrasound: an intelligent image analysis algorithm for diagnosis of malpresentation}},
  booktitle = {Ultrasound},
  year = {2020},
  doi = {10.1002/uog.22267},
}

Loading related content...

Figure