Evaluation of automated tool for two‐dimensional fetal biometry

Abstract

Objective To evaluate whether an automated tool can recognize a structure of interest and measure fetal head circumference (HC), abdominal circumference (AC) and femur length (FL) on frozen two‐dimensional ultrasound images. Methods Ultrasound examinations were performed in 100 singleton pregnancies between 20 and 40 weeks of gestation, ensuring an even distribution throughout gestational age. In each pregnancy, three standard biometric variables (HC, AC, FL) were measured each in three different images obtained for this purpose (i.e. nine independent image acquisitions). An algorithm (Philips Research) was used to detect the structure of interest and automatically place calipers for measurement. Caliper placement was assessed in two ways. First, subjective clinical assessment was performed to determine whether the caliper placement was correct, and caliper position was classified as ‘acceptable for clinical use’, ‘minor adjustment required’ or ‘major adjustment required’. Second, the resulting automatic measurements were compared with manual measurements, taken in real time. Mean difference errors were calculated and expressed as percentages to correct for fetal growth with advancing gestation. Results After exclusion of one pregnancy due to missing images, a total of 891 images (297 for each biometric variable) from 99 pregnancies were analyzed. The algorithm failed to place calipers for the AC in nine images, whereas there were no failures in caliper placement for measurement of HC and FL. On subjective quality assessment of automatic caliper placement, in 475 (53.3%) images position of the calipers was judged to be clinically acceptable and did not require any adjustment, while in 317 (35.6%) and 90 images (10.1%) minor and major adjustments were required, respectively. The mean error between manual and automatic measurement of HC was −0.21 cm corresponding to a percentage error of −0.81% with 95% limits of agreement (LOA) between −3.73% and 2.12%. For AC and FL measurements, the mean error was, respectively, 0.72 cm (percentage error, 2.40%) with LOA between −9.48% and 14.27%, and 0.21 cm (percentage error, 3.76%) with LOA between −8.38% and 15.91%. Conclusions The automated tool identified correctly the biometric variable of interest in 99% of frozen images. The resulting measurements had a high degree of accuracy and compared well with previously published manual‐to‐manual agreement. The measurements exhibited bias, with the automated tool underestimating biometry; this could be overcome by further improvements in the algorithm. Nevertheless, adjustable calipers for manual correction remains a requirement. Copyright © 2018 ISUOG. Published by John Wiley & Sons Ltd.

Using deep‐learning algorithms to classify fetal brain ultrasound images as normal or abnormal

“…It has been reported 14,16,[19][20][21] that images of transventricular and transcerebellar planes can be recognized and biometry performed by CNN-based DL algorithms. While clinical applications of plane detection have been published 22 , verification of abnormality recognition requires very large numbers of images with fetal anomalies.…”

Applications of Advanced Ultrasound Technology in Obstetrics

“…It is a branch of artificial intelligence (AI). In obstetric ultrasonography, the automation of measurements of fetal biometry is a potentially useful tool to increase the reliability and reproducibility of measurements as compared to manual measurements [ 68 ]. In addition, it can reduce scanning time [ 68 ] and work-related fatigue and musculoskeletal disorders [ 69 ].…”

“…Although the inaccuracy for the plane acceptance check for head parameters was 12.9% [ 67 ], such inaccuracy can be corrected by fine-tuning of the caliper placement manually. In another study, manual adjustment of caliper position was not required in about two-thirds of cases for HC and FL measurements, but it was required in more than 80% for the measurement of AC [ 68 ]. Auto measuring AC is more difficult than measuring HC because of the low contrast between the abdomen and surrounding tissues and the large variability in abdominal shape and appearance [ 68 ].…”

Accuracy of automated three‐dimensional ultrasound imaging technique for fetal head biometry

“…Additional techniques developed to increase diagnostic accuracy include fetal neurosonography 7 , magnetic resonance imaging 9 and three-dimensional (3D) ultrasound 10 , which further provide the ability to acquire and process images using AI. A tool for automated recognition and measurements of head circumference (HC), femur length and abdominal circumference on two-dimensional ultrasound images was evaluated recently and showed promising results 11 . Other technologies developed to evaluate fetal intracranial anatomy include Mindray's Smartplanes ® , Samsung's 5D CNS + ® and SonoCNS ® Fetal Brain developed by GE Healthcare (Zipf, Austria).…”