New AI Spots Hidden Heart Defects in Unborn Babies' Ultrasounds

Congenital heart disease (CHD) is the most common birth defect, and a large share of cases slip past prenatal ultrasounds. Part of the problem is that existing AI tools assume someone has already found the important pictures — a technician or a separate program has to isolate the heart views first. A team of researchers removed that step entirely. Their system looks at the whole ultrasound study, hundreds of images at once, and decides from that.

It works in two stages. First, the AI teaches itself what fetal ultrasound images look like by studying unlabeled scans — a bit like learning a language by reading without a dictionary. Then it picks out the frames most likely to show the heart, and a second layer weighs all of them together to produce one verdict for the whole case: likely CHD or not. It also separates critical heart defects from milder ones, and it hands back its top-scoring images so a doctor can see exactly what it reacted to.

The results are striking. On the researchers' own test set, the model scored 0.985 on a measure where 0.5 means random guessing and 1.0 is perfect, and it correctly cleared healthy cases 99% of the time — far ahead of two other AI systems tested alongside it. On scans from a different hospital, though, every model initially performed no better than chance. A technique called CORAL, which adjusts the model without needing new labels, lifted it from 0.513 back up to 0.944.

One honest caveat: this is a preprint, meaning it has not yet been reviewed by other scientists, and it is nowhere near approved for use in hospitals. The external-hospital results show how fragile medical AI can be when scanners and patient populations change — the fix worked here, but it is still a research finding, not a product. A trained sonographer remains essential.