Google Tests AI Ultrasound in Clinics

Google Research reported on October 21, 2026, that a prospective multi-center study found an AI system could estimate gestational age from “blind sweep” ultrasound scans collected by operators without specialized sonography training. The system was tested live on mobile devices in clinics using handheld probes, with minimal fine-tuning. The study compared its estimates with those from expert sonographers, whose standard approach requires trained operators to locate fetal cross-sections and take measurements.

The model was trained on scans from 1,968 participants in North Carolina and Zambia, then fine-tuned using 120 participants in Chicago—about 6% of the original training set—to adapt it to a different probe and extend its prediction range beyond 36 weeks. Researchers tested it prospectively on 192 participants in Chicago and 193 in Nairobi, without using Nairobi data for fine-tuning. The primary cohort included 385 participants between 16 and 36 weeks of pregnancy.

Across Chicago and Nairobi, the AI model’s mean absolute error (average absolute difference from the reference gestational age) was 4.18 days, compared with 4.48 days for expert sonographers; the study found the model non-inferior to the expert approach. Results were similar across settings: 4.10 days in Chicago and 4.27 days in Nairobi, where the model had no local fine-tuning. In exploratory analysis of pregnancies with fetal weight above the 90th percentile, the model outperformed traditional biometry formulas.

The researchers also observed practical issues during deployment, including the value of swappable batteries for keeping devices charged between patients. They said safeguards such as model-confidence estimates could help reject poor-quality scans and communicate uncertainty to clinicians. The study, published in JAMA Network Open, evaluated performance across different patient populations, clinical workflows and ultrasound hardware. More than 250,000 preventable deaths occur worldwide each year from pregnancy-related complications, and two-thirds of people globally lack access to diagnostic medical imaging, according to the World Health Organization estimates cited by Google Research.

Google Research reported on October 21, 2026, that a prospective multi-center study found an AI system could estimate gestational age from “blind sweep” ultrasound scans collected by operators without specialized sonography training. The system was tested live on mobile devices in clinics using handheld probes, with minimal fine-tuning. The study compared its estimates with those from expert sonographers, whose standard approach requires trained operators to locate fetal cross-sections and take measurements.

The model was trained on scans from 1,968 participants in North Carolina and Zambia, then fine-tuned using 120 participants in Chicago—about 6% of the original training set—to adapt it to a different probe and extend its prediction range beyond 36 weeks. Researchers tested it prospectively on 192 participants in Chicago and 193 in Nairobi, without using Nairobi data for fine-tuning. The primary cohort included 385 participants between 16 and 36 weeks of pregnancy.

Across Chicago and Nairobi, the AI model’s mean absolute error (average absolute difference from the reference gestational age) was 4.18 days, compared with 4.48 days for expert sonographers; the study found the model non-inferior to the expert approach. Results were similar across settings: 4.10 days in Chicago and 4.27 days in Nairobi, where the model had no local fine-tuning. In exploratory analysis of pregnancies with fetal weight above the 90th percentile, the model outperformed traditional biometry formulas.

The researchers also observed practical issues during deployment, including the value of swappable batteries for keeping devices charged between patients. They said safeguards such as model-confidence estimates could help reject poor-quality scans and communicate uncertainty to clinicians. The study, published in JAMA Network Open, evaluated performance across different patient populations, clinical workflows and ultrasound hardware. More than 250,000 preventable deaths occur worldwide each year from pregnancy-related complications, and two-thirds of people globally lack access to diagnostic medical imaging, according to the World Health Organization estimates cited by Google Research.