Blind-sweep ultrasound read by AI: novice operators with low-cost probes approach expert gestational-age accuracy in a prospective multicenter study

After an operator without specialized sonography training performs a standardized blind sweep with a low-cost portable probe, the AI system can estimate gestational age directly from the images and run on mobile devices in real clinics. Earlier validation covered only unseen patients from the same retrospective cohort; the conventional approach relies on sonographers trained for years to locate specific fetal cross-sections and take precise biometry measurements, a task that is hard to scale. In the prospective multicenter diagnostic study's primary cohort of 385 participants between 16 and 36 weeks, AI gestational-age estimates were compared with formula-based estimates from the clinically standard expert approach, and the reported accuracy was comparable.

The system shifts gestational-age estimation from expert manual cross-section localization to holistic reading of standardized blind-sweep image sequences. A blind sweep means the operator does not target specific cross-sections but glides the probe across the abdomen in a fixed pattern, moving the localization burden from the person to the model. The model was originally trained on ultrasound data collected in North Carolina and Zambia with Sonosite, GE and Butterfly Networks commercial probes, then paired with the Clarius C3 commercial handheld transabdominal probe; tuning used only a pilot dataset of 120 Chicago participants, about 6% of the original training set, and no Nairobi site data was used for fine-tuning, keeping the Kenyan validation a strictly independent external test.

In cases of fetal weight extremes above the 90th percentile, the model outperformed traditional biometry formulas. This is an exploratory metric beyond accuracy, pointing to robustness against variation in fetal size. The result comes from exploratory analysis of complex scenarios in the same prospective multicenter study, which also included multiple gestation pregnancies that can present tracking and measurement challenges for both human operators and automated systems.

The study evaluated the system in a prospective multicenter setting across two target environments, Chicago and Nairobi, with a primary cohort of 385 pregnant participants between 16 and 36 weeks, operators without specialized sonography training, the Clarius C3 handheld probe, and comparison against formula-based estimates from expert operators using traditional cart-based ultrasound.

Tuning used only a pilot dataset of 120 Chicago participants, about 6% of the original training set, and no Nairobi site data was used for fine-tuning; the advantage in fetal weight extremes is an exploratory metric. The research team notes that deploying such systems will also require building deep trust with clinicians and patients, and that clear safeguards such as model confidence estimates could reject poor-quality sweeps and convey model uncertainty to the clinician user.