AI in Obstetrics: Automated Fetal Biometry and CTG Interpretation

For decades, imaging and fetal monitoring in obstetric practice have been the work of an experienced eye and hand. The sonographer angling the probe to the right plane, freezing the image, placing the calipers, and recording the measurement — every link of this chain is dependent on operator skill, fatigue, and inter-observer variability. Over the past three years, deep-learning–based systems have entered this chain with claims of improving both speed and reproducibility, and have moved to the center of the clinical research agenda. In this article I address, with current evidence, where artificial intelligence actually stands in automated fetal biometry, anomaly-screening support, and cardiotocography (CTG/EFM) interpretation.

Automation in Fetal Biometry: Handing Measurement to the Machine

The backbone of the second-trimester scan is biometry: head circumference (HC), biparietal diameter (BPD), abdominal circumference (AC), and femur length (FL). These four measurements underpin both fetal growth assessment and gestational-age dating. In the traditional workflow, the sonographer pauses scanning for each measurement, captures the plane, and places calipers manually.

A notable study published in npj Digital Medicine in 2024 turned the paradigm on its head: instead of "capture the right frame and measure," the approach gathers automated measurements from every frame of the scan. The system classifies each frame of the 20-week scan with a neural network, measures biometry in every frame in which appropriate anatomy is visible, and uses a Bayesian method to estimate the true value from hundreds of measurements while probabilistically discarding outliers. In a retrospective experiment with 1,457 records and roughly 48 million frames, the estimated biometrics matched real-time manual measurements at human-level performance.

AI in Anomaly Screening: The PROMETHEUS Randomized Trial

Automated measurement alone is not enough; the real question is whether AI improves the workflow of anomaly screening without harming diagnostic accuracy. The most concrete answer to that question to date came from the PROMETHEUS randomized controlled trial presented in 2024.

In the trial, 78 pregnant women (26 carrying a fetus with congenital heart disease) were scanned both with the standard method and with the AI-assisted method; sonographers were randomized. In AI-assisted scans, the model recognized and captured 13 standard planes and measured four biometric parameters. The result was meaningful: AI assistance produced a significant time saving and a reduction in sonographer cognitive load in routine fetal anomaly screening, without a drop in diagnostic performance. The study was presented at the 34th ISUOG World Congress (Budapest, September 2024).

PROMETHEUS's contribution is not the claim that "AI finds anomalies better than humans." The message is more realistic and more useful: AI preserves the experienced sonographer's accuracy while speeding up the work and lowering mental load. The sample is small (78 pregnancies); larger validations are expected.

The ISUOG Framework: No Automation Without Standardization

The International Society of Ultrasound in Obstetrics and Gynecology (ISUOG) sets standards in this field through its practical guidelines on fetal biometry and growth assessment (2019) and on the conduct of the routine mid-trimester scan (2022 updated version). The value of AI models hinges greatly on whether they were trained on standards-compliant, correctly labeled planes. ISUOG has been bringing ultrasound + AI in obstetrics onto its agenda through dedicated educational events — the professional framework is not rejecting the technology but seeking a standardized footing for it.

This point is critical: a biometry AI's clinical validity is limited to the population, device, and plane definitions on which it was trained. Guideline alignment is the healthiest brake against the assumption that a model "works everywhere."

AI in CTG / EFM Interpretation: The Most Awaited, Hardest Area

Cardiotocography (CTG) is routinely used to evaluate fetal well-being during labor; its interpretation, however, is highly subjective. By the nature of visual assessment, intra- and inter-observer variability is high, and the high false-positive rate of continuous intrapartum monitoring, with limited improvement in fetal outcomes, is a long-standing issue. This is precisely why CTG is one of the most hoped-for — and one of the hardest — application areas for AI.

What do current models do?

A study published in American Journal of Obstetrics & Gynecology in 2025 (and presented at the SMFM meeting in February 2024) used intrapartum electronic fetal heart rate monitoring with deep learning to predict acidemia at birth. Different groups are building models with CNN, Transformer, LSTM, and CfC architectures to predict fetal hypoxia from CTG traces; national multicenter studies are comparing AI to humans.

Table: Maturity of Three Application Areas in Obstetric AI

Two large barriers

• Proving superiority over the human expert: CTG's historical issue is a high false-positive rate. For an AI to add clinical value, it must not only "predict correctly" but reduce unnecessary intervention while not missing poor outcomes — a bar that is hard to clear.
• Explainability: In the labor ward, where seconds matter, clinical trust will not form if the model cannot show the clinician why it raised an alert. Real-world use of current models is limited for exactly these two reasons.

The fact that previous decision-support systems for CTG (such as the INFANT trial) failed to meaningfully improve perinatal outcomes underscored that adding technology alone is not enough. Newer deep-learning models, mindful of this lesson, focus on hard endpoints such as acidemia/hypoxia — but prospective, outcome-focused evidence is still required for them to enter clinical routine.

A Practical View from the Field

As an OB/GYN evaluating these technologies, I ask three questions: (1) Does the model resemble my patient population and my device? (2) Can I see when it gets it wrong (explainability)? (3) Does the evidence say "accuracy preserved" or "outcomes improved"? The honest picture today is that automated biometry and AI-assisted anomaly screening have demonstrably improved speed and reproducibility; CTG interpretation remains promising but has not clearly moved beyond the "assistant" role. AI is not replacing the experienced clinician; it is supporting the measuring hand and the tiring eye.