A doctor stares at a gray-white smudge on the screen. It's the face of an unborn child. One shadow in the right place, and you can plan surgery in time. But not everyone sees that shadow: cleft lip or palate is rare, and a young specialist might go years without ever spotting one.
The neural network was trained on 45,139 ultrasound images from 9,215 fetuses across 22 hospitals. Now it finds clefts with a sensitivity above 93% and an accuracy above 95%. That's the level of a senior radiologist. When the system was brought in as a "second pair of eyes," junior doctors caught 6% more cases.
In a pilot study, 24 people, from students to physicians, trained on images alongside the model. It turned out that AI accelerates the acquisition of experience for rare diagnoses. It's like a simulator that shows you where to look and explains what you're seeing.
This dual benefit, accurate diagnosis and training, is especially valuable where experienced specialists are scarce.
An ultrasound machine sends sound waves into the body, and they bounce back as echoes from tissues of different densities. From those echoes, the doctor assembles an image, much like a bat "sees" obstacles in the dark.