MIT’s E(3)-Pose tracks fetal movement to improve MRI image acquisition

The research model builds geometric symmetry into its architecture so a scanner can follow a moving fetal head.

Researchers at MIT, Boston Children’s Hospital and Inria have developed E(3)-Pose, a deep-learning model that tracks the position of a fetus’s head during MRI acquisition. MIT described the work on 19 October, following presentation of the paper at the European Conference on Computer Vision.

The technical idea is to incorporate geometric structure into the model. Left-right symmetry helps the network identify orientation, allowing acquisition to respond to movement instead of treating every captured slice as though the subject had stayed still.

The current model was trained primarily on third-trimester fetuses. Extending it to earlier pregnancy is a stated research direction rather than an established capability.

Moving the intelligence into acquisition

An MRI volume is assembled from two-dimensional slices. When the subject moves between them, a set of individually useful images may fail to form a coherent three-dimensional picture. Tracking motion changes the question from how to repair a completed image to how to acquire the next useful image.

That is a significant systems distinction. The model becomes part of a feedback loop involving the patient, scanner and acquisition sequence. A good orientation estimate has to arrive in time to influence the scan, while the resulting slice must add information that the final volume needs.

For engineers, the complete workflow therefore includes more than model inference. Coordinates have to align between the network and scanner, timing has to fit the acquisition process, and the system has to handle moments when a new estimate is unavailable. Those interfaces determine whether an algorithm can become a useful part of an imaging instrument.

Geometry provides a useful starting structure

Building symmetry into a network supplies information that would otherwise have to be learned from examples. In this case, the relevant structure is closely connected to the physical task: determining how a head is oriented in three-dimensional space.

The broader attraction of this approach is efficiency of reasoning. A model that already represents meaningful geometric relationships can devote more of its learning to the variation within those relationships. It is similar to choosing a coordinate system that makes a problem easier before beginning the calculation.

That does not remove the need for representative data. It gives the data a more structured role. Researchers can ask which movements, stages of development and scan conditions remain difficult, then design acquisition and training work around those gaps.

A project linking computer science and scanner operation

The development described by MIT included work with radiology researchers, scanner sequences and pregnant volunteers. That combination is important because the output is intended to support image interpretation by a radiologist, rather than produce a diagnosis independently.

For a hospital evaluating acquisition assistance, the meaningful outcome is a usable examination: the required anatomy is covered, image quality supports the clinical question, and the patient’s time in the scanner is managed effectively. Motion tracking is a means of improving that examination.

E(3)-Pose is an example of AI moving closer to the point where data is created. Instead of analysing whatever an instrument happened to collect, the system helps shape the collection itself. Its next development steps will show how that idea performs across a wider range of fetal motion and imaging conditions.

Sources

- [Why teaching AI about symmetry could make prenatal diagnosis more reliable](https://news.mit.edu/2026/teaching-ai-about-symmetry-could-make-prenatal-diagnosis-more-reliable-1019) — MIT News · 2026-10-19