Estimating fetal weight from ultrasound — AI vs. the standard Hadlock method — a clinical trial (ClinicalTrials.gov)

During pregnancy, ultrasound is used to measure features such as the fetal head, abdominal circumference, and femur length, which are fed into a formula to produce an "estimated fetal weight." This estimate helps detect growth restriction or a large-for-dates fetus and informs decisions about the timing and mode of delivery. The Hadlock formula, which plugs these measurements into an equation, has long been the standard, but its estimates carry a margin of error that can widen with maternal body habitus or pregnancy complications. Analyzing the ultrasound image itself with AI could, in principle, yield a steadier estimate without relying solely on a few measurement points.

What distinguishes this study is that it does not simply claim AI "gives better numbers." It compares AI models against the traditional formula on equal footing and goes further to examine whether the AI is systematically off for particular groups (fairness). AI tends to inherit biases in its training data, and if accuracy varies by body size, parity, gestational age, maternal age, fetal sex, or the presence of preeclampsia, some patients could be disadvantaged. A design that confronts such bias head-on is an important precursor to bringing medical AI into routine care.

Estimating fetal weight is one of the most frequently performed assessments in everyday obstetric practice, so any improvement in accuracy could ripple across perinatal management. Comparing image-derived AI estimates with an established formula — and evaluating fairness alongside accuracy — stands as one example of a methodology relevant to medical AI broadly: testing imaging AI on both "accuracy" and "fairness."