NederlandsAI that knows a head is symmetrical tracks the moving head of an unborn baby during an MRI scan
An unborn baby moves during an MRI scan, so images of the brain fail or come out skewed. Researchers at MIT and Boston Children's Hospital built an AI model that quickly tracks the position of the head. The secret: the model already knows that a head is the same on the left and the right.
Sometimes doctors want to get a good look at the brain of an unborn baby, for example when an ultrasound shows something unusual. An MRI scan then gives much sharper images. But there is a problem: a baby in the womb does not lie still.
An MRI scanner makes the image of the brain in slices, a bit like a loaf of bread being cut into slices. The computer then builds a 3D image from all those flat slices. If the baby moves in between, the slices are taken at different angles. Then parts are missing or lie skewed relative to each other. As a result, a radiologist can miss an abnormality or judge it wrongly. Often the scan then has to be repeated, which is long and unpleasant for the pregnant mother.
Watching along while the baby moves
Researchers at the American MIT, Boston Children's Hospital, Harvard Medical School and the French research institute Inria are working on a solution. The idea: in between, the scanner keeps making a very fast, rough 3D recording. An AI model calculates from it how the baby's head is lying at that moment. The scanner then adjusts the angle of the next slice, so that the slices fit neatly together.
That has to happen at lightning speed. The new model, called E(3)-Pose, needs about 0.3 seconds per recording, according to the researchers.
The trick: symmetry is already built in
The difficult part is that those fast recordings are rough and blurry. A head is also almost symmetrical: the left and right halves look very much alike. Earlier AI models got confused by this. They sometimes could not tell which side was left and which was right, and then estimated the position of the head quite wrongly.
The researchers therefore built the symmetry directly into the model. The model does not have to learn from examples that a head is the same on the left and right: it already knows. The model also knows that a head stays the same head when it turns. Compare it with someone who has to follow a football: if you already know the ball is round, you only need to watch where it rolls.
How well does it work?
The researchers tested the model on scans of more than a hundred pregnant volunteers and on clinical scans of younger babies, between 18 and 23 weeks of pregnancy. In the test with real series of fast recordings, the model was on average about 9 degrees off when estimating the rotation of the head. The best earlier method was more than twice as far off. On the clinical scans, too, the new model was the most accurate.
The researchers presented their work at the 2026 European Conference on Computer Vision (ECCV). Their code is public.
Caveats and next steps
The model was trained mainly on babies in the third trimester, the last three months of pregnancy. First author Ramya Muthukrishnan now wants to extend it to younger babies. They move more, and it is precisely for them that an early diagnosis matters.
Building the technique into a real scanner also turned out to be a lot of work: according to the researchers, the scanner crashed several times during the trials. Most tests so far have been done on scans made earlier. The next step is to let the model steer during real scans.
What does this mean for the road to the singularity?
This research shows a smart lesson for AI in general. You do not have to make a model learn everything from examples. What you already know for certain about nature, such as symmetry, you can build in directly. The model then needs fewer examples and makes fewer strange mistakes. That is especially valuable in medicine, where good training data are scarce.
The researchers also see uses outside the hospital, for example in self-driving cars that have to track objects in 3D with limited sensors. Such cross-fertilisation between fields is becoming more common: an idea from mathematics helps in pregnancy care, and can then go further elsewhere. In time, this kind of AI can help make every scan succeed in one go, with a faster and better diagnosis for the youngest patients.
What is the technological singularity? Read the explanation →
