Whole-examination AI estimation of fetal biometrics from 20-week ultrasound scans

TL;DR

The paper presents a paradigm shift for fetal biometrics by automatically classifying every frame of a 20-week ultrasound video, applying per-frame biometric estimation, and aggregating these measurements with a Bayesian framework to obtain a true biometric value and credible intervals. The approach eliminates operator-dependent plane selection and caliper placement biases, enabling real-time, progressively refined estimates with quantified uncertainty. Across a large retrospective dataset, single-frame estimates show small biases but moderate frame-level variability, while whole-scan aggregation achieves human-level agreement and robust test–retest reliability, even under domain shifts. The work demonstrates practical benefits, including potential reductions in scan time and improved reproducibility, and lays groundwork for prospective trials to assess impact on detection of fetal abnormalities.

Abstract

The current approach to fetal anomaly screening is based on biometric measurements derived from individually selected ultrasound images. In this paper, we introduce a paradigm shift that attains human-level performance in biometric measurement by aggregating automatically extracted biometrics from every frame across an entire scan, with no need for operator intervention. We use a convolutional neural network to classify each frame of an ultrasound video recording. We then measure fetal biometrics in every frame where appropriate anatomy is visible. We use a Bayesian method to estimate the true value of each biometric from a large number of measurements and probabilistically reject outliers. We performed a retrospective experiment on 1457 recordings (comprising 48 million frames) of 20-week ultrasound scans, estimated fetal biometrics in those scans and compared our estimates to the measurements sonographers took during the scan. Our method achieves human-level performance in estimating fetal biometrics and estimates well-calibrated credible intervals in which the true biometric value is expected to lie.

Figures (16)

• Figure 1: Example frames with annotations and calipers. Shown are (a) a Brain-TV image, with annotations for head circumference, biparietal diameter and posterior ventricle, (b) a femur image, with the femur length measured and annotated. These frames were acquired and saved by a sonographer.
• Figure 2: CaliperNet example heatmap outputs for (a) a femur image, and (b) a brain-CB image with labels for the cerebellum, cisterna magna and nuchal fold.
• Figure 3: (a) Part of a scale bar, cropped from an image frame. Two lines are overlaid upon it. (b) The pixel values along those two scan lines for the entire frame. (c) The autocorrelation of the pixel values of the blue line after it has been processed with a high-pass filter.
• Figure 4: Images used for biometric training showing (a) femur, and (b) head views. The second row shows the training labels and the third row shows the CNN output overlaid on the image. For the head image, the HC output is shown in green and the BPD output in pink.
• Figure 5: The pipeline which each frame in an US scan feed is processed by to estimate biometrics.