Fetal gestational age estimation using artificial intelligence on non-targeted ultrasound images and video

Abstract

We developed a deep learning model trained on over two million ultrasound images from 78,531 pregnancies from Australia, India, and the UK to estimate gestational age (GA) directly from any fetal ultrasound image, regardless of orientation. The model outputs both a GA estimate and an uncertainty value based on image quality. Independent validation on 36,762 ultrasound images from 742 fetuses showed a mean absolute error (MAE) of 1.7 days at 14-18 weeks and 2.8 days at 18-24 weeks, significantly outperforming traditional biometry (p < 0.001). In video analysis, the model achieved a median prediction time of 24 s and an MAE below 3 days across all trimesters. Performance was consistent across maternal body mass index (BMI) categories and geographic settings. This AI-based GA estimation method matches or exceeds gold-standard fetal biometry, reduces reliance on highly skilled sonologists, and offers the potential to improve access to prenatal care in resource-limited and underserved settings globally.

Fig. 1. Kalman filtering process for generating predictions on video.

In this figure, μ₁, … ,μ_n and σ₁, … ,σ_n denote sequences of mean and variance predictions from the model. p₁, … ,p_n and q₁, … ,q_n respectively denote sequences of state variables for the Kalman filter process. Σ is used to denote a threshold value on the variance predictions, below which no update is made.

Fig. 2. Predicted versus actual gestational age.

Scatter plot comparing the gestational age (GA) predicted by the AI model with the actual GA based on gold standard crown–rump length dating (plus time elapsed). Each point represents a prediction for a single case. The data points cluster closely and uniformly around the diagonal line, demonstrating high concordance between predicted and actual values.

Fig. 3. Bland–Altman plot of predicted versus actual gestational age.

This plot illustrates the agreement between the AI-predicted gestational age and the actual (gold standard) gestational age.

Fig. 4. Cumulative distribution of the time taken to generate gestational age prediction.

The cumulative distribution of time (in seconds) required by the AI model to produce a sufficiently confident gestational age (GA) estimate from video data. The x-axis represents elapsed time since the start of the video, and the y-axis indicates the proportion of cases for which the model had already generated a prediction. The model typically produces estimates rapidly, with a median prediction time of 24 s, and over 95% of predictions are completed within 60 s. This highlights the efficiency of the model in real-time scanning contexts.