Research AI tracks a moving fetal brain during MRI scans

A fetal MRI is a long wait for a mother and a hard scan for a radiologist. The mother can spend an hour inside the machine and still quit halfway, and the doctor may end up without the full, sharp image needed to follow how the baby's brain is developing. Ramya Muthukrishnan, an MIT Ph.D. student, saw that frustration firsthand when she shadowed a radiologist at Boston Children's Hospital at the start of her doctorate.

The problem is mechanical. A fetal MRI is assembled from a series of 2D images, or slices, stacked into one 3D picture of the head. If the baby moves, each slice can end up taken from a different angle, leaving gaps in coverage and unusual slice orientations. As Muthukrishnan puts it, "If you have coverage gaps, you can't see critical regions, and you need to be able to see well to detect abnormal development."

The model she helped build, E(3)-Pose, is designed to follow that motion. Its distinguishing feature is the use of physics rather than data alone: the left-right symmetry of the fetal head is embedded in the neural network's architecture, instead of being left for the model to learn from examples. This makes it easier for the model to identify the orientation of the brain while it moves. The work is led by Muthukrishnan in the lab of Polina Golland, with co-authors at Harvard Medical School and Boston Children's, and with Benjamin Billot of Inria, who developed the initial method. Borjan Gagoski developed an MRI sequence that acquires additional scans to help train the model, and Aryn Lee recruited the pregnant volunteers.

The team's own account is that the project was harder than the theory. Muthukrishnan, who came from computer science, says she repeatedly crashed the scanner when she moved the method from training on a GPU to running on the hardware. The paper was presented at the 2026 European Conference on Computer Vision and is available as an arXiv preprint, which means it is publicly posted but has not yet been through a clinical study.

For parents and clinicians, the promise is a scan that is both more complete and easier to sit through. Grant, who directs the Fetal Neonatal Neuroimaging and Developmental Science Center at Boston Children's, expects the approach to automate fetal brain MRI, replacing a slow manual process. The researchers also say the method could help with other hard-to-image anatomy, and even with autonomous driving. The sources do not give a timeline for hospital use, and the results still need to be confirmed in routine clinical conditions before any change in how fetal MRIs are done.