Researchers at Google are exploring AI ultrasound to help more pregnant women get the prenatal scans they need.

The work was a research study with Northwestern and Jacaranda Health. In it, healthcare workers were trained to perform simple ultrasounds. Machine learning models then interpreted the results.

Googlers Angelica Willis and Dr. Nichole Young-Lin shared details of the research in an “Ask a Scientist” interview.

Why ultrasounds matter in pregnancy

Ultrasounds give expecting parents a lot of information. They can hear a heartbeat, count fingers and toes, and check how many babies there are. They can also see how each baby is positioned.

Doctors use them for clinical questions too. Is the baby growing appropriately? Are there anomalies?

Early scans help estimate gestational age. That tells doctors how far along the pregnancy is and when the baby is due.

Dr. Young-Lin is the in-house OB/GYN at Google. She explained why this timeline matters. Screenings, treatment, management and delivery planning all depend on it.

Period history is not always reliable. Many people have irregular periods or do not track them.

Errors can have serious effects. A baby believed to be 37 weeks old might actually be 34 weeks old, with less mature lungs. Those babies need very different levels of care.

The access gap

Not every pregnancy gets this level of care. About two-thirds of people worldwide lack ready access to diagnostic imaging, such as ultrasounds and X-rays. There are also not enough sonographers.

Equipment is part of the problem. Traditional ultrasound machines are bulky and expensive. In remote clinics, there is often no one to fix them when they break. They also need stable electricity.

Handheld devices help. They are smaller and cheaper, and many run on batteries. That makes them more portable and accessible.

Training is another barrier. According to Willis, it still takes two years of intense hands-on training to become a sonographer.

How blind sweep scans work

The study took a different approach. It used a “blind sweep” ultrasound.

In a traditional scan, a sonographer must move the probe precisely to capture specific measurements. That is hard to do well.

In a blind sweep, the operator moves the probe across the abdomen in a predefined pattern. Healthcare workers who have never performed an ultrasound can learn the method in eight hours.

The operators collect videos of the baby. An AI model then analyzes them. It can tell the operator when a sweep needs to be repeated. It can also estimate gestational age and indicate fetal presentation, meaning the baby’s position.

Importantly, all processing happens on the device. It does not need an electricity supply or Wi-Fi. The operator and the mother still receive the insights they need to prepare for birth.

Study results

The team partnered with Jacaranda Health and Northwestern Medicine. The study included 1,000 mothers in Nairobi, Kenya, and 1,000 more in Chicago.

The machine learning model detected gestational age and fetal presentation as accurately as a trained sonographer. This points to better information for women in under-resourced areas ahead of birth.

What could come next

Willis says the research shows that expert-level care can reach frontline community health workers. Dr. Young-Lin says the results show that AI ultrasound is a viable application of AI in maternal healthcare. It worked across different patient populations and healthcare settings.

She also sees wider uses. Ultrasound can show how much someone is bleeding after an accident or injury, for example. She adds that there is much potential for AI to improve maternal health and healthcare across the board.

For now, the study offers a clear example. AI ultrasound could help bring prenatal imaging to women who have struggled to get it.

Related post

View all