Fetal brain MRI can be difficult to interpret when a fetus moves during a scan. Each image slice may capture the head from a different angle, leaving gaps in the final three-dimensional picture. Researchers at MIT, Boston Children’s Hospital, and Inria have developed E(3)-Pose, a model that uses a built-in understanding of symmetry to help track that motion.
Why movement complicates fetal MRI
A routine prenatal ultrasound may flag a possible brain abnormality for further examination with magnetic resonance imaging, or MRI. But imperfect imaging can lead to missed or incorrect diagnoses, and motion during the MRI can make it harder to produce a complete view of the fetal brain.
An MRI scan gathers a series of two-dimensional images, called slices. When those planes are combined, they form a higher-resolution, three-dimensional image of the head. If the fetus shifts, however, the slices may be taken from different angles. The resulting image can contain gaps or unusual orientations.
Those gaps matter because radiologists need a clear view of fine details to assess development. If a critical region is missing from the reconstructed image, a physician may have less information available when reviewing the scan for a possible pathology.
Putting symmetry into the model
E(3)-Pose is a deep learning model designed to help image the fetal head while it is moving. Its approach builds a principle of physics into the neural network: the left-right symmetry of the fetal head. The researchers say this gives the model a useful cue for identifying the brain’s orientation.
That design differs from relying only on a model to learn patterns from data. By incorporating symmetry into the architecture itself, the team gives the system a structural guide as it works with images taken from changing positions. The aim is to help the scanner capture the information needed for a clearer, more complete image that a radiologist can examine.
The work was presented at the 2026 European Conference on Computer Vision. First author Ramya Muthukrishnan is an MIT PhD student in the lab of Polina Golland, a principal investigator at the MIT Abdul Latif Jameel Clinic for Machine Learning in Health and the MIT Computer Science and Artificial Intelligence Laboratory.
From a technical problem to a clinical need
Developing the model also meant working with MRI hardware. Muthukrishnan entered the project with a computer science background and had to learn how to implement the method on a scanner. She described the practical challenges of getting hardware to work as different from training a model on data using a GPU.
Her motivation came partly from observing fetal MRI scans at Boston Children’s Hospital. She saw a pregnant patient become uncomfortable during a lengthy scan and stop before it was finished, while the radiologist was left without a complete, high-quality MRI for monitoring fetal development. That experience helped define the problem she wanted her doctoral work to address.
Several collaborators contributed to the project. Borjan Gagoski developed a sequence that acquires additional MRI scans to support training E(3)-Pose and helped troubleshoot the scanner. Aryn Lee recruited pregnant volunteers and assisted with MRI acquisition. Senior authors guiding the work included Elfar Adalsteinsson, Ellen Grant, and Golland; Benjamin Billot helped Muthukrishnan get started with the initial method he developed.
What the team hopes to improve
The model is primarily trained on third-trimester fetuses. Muthukrishnan hopes to extend its usefulness to second-trimester or younger fetuses, which are smaller and move more because there is more space in the uterus. The team’s stated aim is to make earlier imaging support possible, when detecting brain abnormalities could make intervention easier.
The potential use of E(3)-Pose may also reach beyond fetal MRI. The researchers say it could help with imaging other anatomy that is difficult to capture, and might apply to autonomous driving, where limited sensors must track three-dimensional objects using noisy data.
For prenatal care, the central promise is a better path from moving anatomy to an image a physician can assess. E(3)-Pose is intended to help the scanner account for fetal motion and produce higher-quality imaging, while the radiologist remains the one examining the resulting scan.