MIT details AI that uses head symmetry to track movement during fetal MRI

A moving fetal head can leave gaps in an MRI scan, obscuring details needed to assess brain development. In an October 19 research account, MIT described E(3)-Pose, an AI model that builds head symmetry into its design to track movement during imaging. The goal is more complete scans—not an AI diagnosis.

Researchers at MIT, Boston Children’s Hospital, and Inria developed the model. Their work was presented at the 2026 European Conference on Computer Vision. The underlying paper first appeared on arXiv in December 2025 and was revised in June 2026; MIT’s account describes the project’s imaging approach, scanner implementation, and remaining challenges.

Learning symmetry versus building it in

Fetal MRI captures a series of two-dimensional images, or slices, that together form a detailed three-dimensional view of the head. If the fetus moves between slices, the images can come from different angles. The resulting gaps and unusual orientations can prevent a radiologist from seeing regions needed to identify abnormal brain development.

The model addresses that problem by estimating the head’s position and orientation. Rather than relying on training examples alone to teach it the relevant geometry, E(3)-Pose embeds the head’s left-right symmetry in its architecture. It also explicitly models how its estimates should change when the head rotates—a property called rotation equivariance.

Symmetry is both useful information and a source of ambiguity. In their paper, the researchers explain that anatomical symmetry, low resolution, noise, and imaging artifacts make existing pose-estimation methods struggle with clinical scans. Their approach represents symmetry directly instead of leaving the network to infer it entirely from the data.

From a prediction to the next MRI slice

The proposed scanning process uses a rapid, low-resolution 3D MRI volume acquired before each diagnostic 2D slice. E(3)-Pose estimates the head’s position and orientation from that volume. Those estimates are intended to let the scanner automatically adjust where it captures the next slice as the fetus moves, rather than accept a mismatched view.

Getting that process onto an MRI scanner brought challenges beyond model training. First author Ramya Muthukrishnan, an MIT PhD student, described repeated scanner crashes during implementation:

I was not prepared for how many times I would crash the scanner because putting things on hardware and getting it to work on hardware is a different beast than training a model to run on data on a GPU.

Ramya Muthukrishnan, first author, speaking to MIT News

Co-author Borjan Gagoski, a researcher at Boston Children’s Hospital, developed a scanning sequence that collected additional MRI scans for training. He also helped troubleshoot the scanner. Research assistant Aryn Lee recruited pregnant volunteers and helped acquire the scans, connecting the model-development work to data collection on the imaging hardware.

Tracking accuracy is not diagnostic proof

The researchers report experiments on public and clinical fetal MRI datasets, finding stronger robustness and generalization than comparison methods. They also report state-of-the-art head-pose accuracy on clinical MRI volumes. The authors frame those results as support for future clinical use, not proof of better diagnoses.

The clinical ambition is to replace a slow, manual scanning process with faster, more comfortable exams and higher-quality images, according to senior author Ellen Grant of Boston Children’s. That prospect reflects a problem Muthukrishnan observed while shadowing a radiologist: a pregnant patient stopped an uncomfortable scan before a complete, high-quality MRI could be obtained.

An important next target is earlier pregnancy. E(3)-Pose was primarily trained on third-trimester fetuses, and Muthukrishnan hopes to improve it for second-trimester or younger fetuses. She says smaller fetuses move more because they have more room in the uterus. Those scans are especially important to her because earlier detection of brain abnormalities can make intervention easier.