source: Google Research: Empowering frontline maternal care with AI-assisted ultrasound

level: research

google research and partners tested an ai system that reads blind sweep ultrasounds, where a novice glides a low-cost probe across the abdomen. the model estimated gestational age from these sweeps. it was compared to expert sonographer biometry, the standard of care. the study ran prospectively in chicago and nairobi. the ai achieved a mean absolute error of 4.18 days versus 4.48 days for experts, showing non-inferiority.

the model was trained on 1,968 participants from north carolina and zambia. it was fine-tuned on only 120 chicago participants, about 6 percent of the original data. no nairobi data was used for tuning. in chicago, the error was 4.10 days. in nairobi, it was 4.27 days. the ai also handled fetal weight extremes better than traditional formulas. the study included 385 participants between 16 and 36 weeks of pregnancy.

the work addresses a gap where two-thirds of people globally lack ultrasound access. traditional sonography training takes years and is costly. blind sweep scanning needs minimal training. the ai ran live on mobile devices with a clarius c3 probe. researchers noted practical needs like swappable batteries and model confidence estimates to reject poor sweeps. this could expand prenatal screening in low-resource settings.

why it matters: this ai could let community health workers provide accurate pregnancy dating without years of sonography training, expanding prenatal care in underserved areas.


source: Google Research: Empowering frontline maternal care with AI-assisted ultrasound