Fetal MRI stops fighting motion: symmetry built into the network aligns slices as the head moves
Synopsis
When a fetus moves during an MRI scan, the 2D slices are acquired at shifting angles, leaving gaps in the final 3D image that can cause a radiologist to miss or misread a brain abnormality. Researchers at MIT, Boston Children's Hospital and Inria built E(3)-Pose, a deep learning model that tracks the fetal head during imaging to help capture a high-resolution image a radiologist can examine. The model embeds left-right symmetry of the fetal head directly in its architecture rather than learning physics only from data, and it is trained mainly on third-trimester fetuses.
Interpretation
Fetal MRI can now track head pose while the fetus moves, realigning 2D slices taken at scattered angles and reducing coverage gaps in the 3D reconstruction. Previously, motion meant each 2D slice came from a different angle, producing gaps and unusual slice orientations that could lead a physician to miss or misdiagnose a fetal brain pathology. The team acquired fetal MRI scans at Boston Children's Hospital and developed a sequence that collects additional scans to train E(3)-Pose; the work was presented at the 2026 European Conference on Computer Vision.
The model tracks motion because physics is written into the neural network, not merely inferred from data. Instead of training the model to learn physics from data alone, left-right symmetry of the fetal head is embedded in the architecture, making the brain's orientation easier to identify. E(3)-Pose was developed and trained by researchers at MIT, Boston Children's Hospital and Inria, led by first author Ramya Muthukrishnan, an MIT PhD student, with senior authors including Polina Golland, Elfar Adalsteinsson and Ellen Grant.
The demonstrated scope centers on third-trimester fetuses, with earlier gestation framed as the next target. The model is primarily trained on third-trimester fetuses, whereas second-trimester fetuses are smaller and move around more because there is more space in the uterus. Muthukrishnan says she hopes to improve E(3)-Pose to help with second-trimester or younger fetuses, reasoning that earlier detection of brain abnormalities makes intervention easier.
Perspective
The report is for researchers and engineers working on medical imaging and deep learning, covering E(3)-Pose's motivation, its physics-prior architecture and its clinical acquisition setting. The model is primarily trained on third-trimester fetuses, and the authors explicitly aim to extend it to second-trimester and younger fetuses. The material also suggests transfer to other hard-to-image anatomy and to 3D object tracking with limited sensors, but that is a prospect rather than a validated result.
The material gives no quantitative tracking accuracy, reconstruction quality or diagnostic accuracy, so the size of any improvement cannot be judged from it. The model is trained mainly on third-trimester fetuses, and whether it works for smaller, more active second-trimester fetuses remains to be tested. The cross-domain applications the authors mention are outlook, not experimental results.