Teach an AI symmetry instead of making it guess

What is it?

An MRI scan takes pictures of the inside of the body, one thin slice at a time. If a baby moves during the scan, the slices no longer line up. Doctors may have to scan again.

A team at MIT, Boston Children’s Hospital and Inria built an AI called E(3)-Pose. It looks at a quick, low-detail 3D scan of the baby’s head and works out how the head is turned and tilted. The aim is to adjust the next slice to match.

MIT News reported the work on 19 October 2026. The paper was first posted on arXiv in December 2025 and last updated on 25 June 2026.

What did they do differently?

A head is almost the same on the left and the right. In a blurry scan, that makes it hard to tell which side is which. A normal AI sees many examples and tries to learn this on its own.

This team did not leave it to chance. They built the rule into the AI’s design: the left side matches the right side. The paper says the AI captures this symmetry “by construction”.

How well did it work?

On the authors’ own tests, E(3)-Pose did better than other methods on scans that look like real hospital scans, including quick low-detail ones. Each estimate takes a fraction of a second.

0.3 s

time for each estimate, including one extra step, for the fast methods compared

Why should you care?

You can use the same idea. If you know a rule about your problem, you can often build it in. The AI then needs fewer examples and makes fewer silly mistakes.

The paper says the method may help with other jobs too, where things are symmetric, there is little training data, and the sensors are noisy.

What’s the catch?

• This is research. Nothing in the paper or the MIT News story says it is in use in a hospital.
• The test with real scan-style data came from one hospital and one scanner.
• Real heads are not perfectly symmetric. Large one-sided differences could break the rule. The authors say they have not tested real cases of that.
• MIT News says the model learned mostly from older babies and must improve for younger ones, who move more.

Try it

Cut a butterfly from folded paper. Cover the right wing and draw the left one by looking at the right. You used a rule, not a guess. Now ask: what rule do you already know about a problem you are solving?

Sources

• Equivariant Symmetry-Aware Head Pose Estimation for Fetal MRI, Muthukrishnan, Gagoski, Lee, Grant, Adalsteinsson, Billot and Golland, arXiv, version of 25 June 2026
• Why teaching AI about symmetry could make prenatal diagnosis more reliable, MIT News, 19 October 2026