When a routine prenatal ultrasound suggests a possible brain abnormality, the next step is often an MRI. The trouble is that a fetus moves. An MRI scan of the head is built from a stack of flat 2D slices, and if the head shifts between them, slices end up at odd angles with gaps between them, which can lead a radiologist to miss or misread a problem. Researchers at MIT, Boston Children's Hospital and the French institute Inria have built a model meant to keep up with that movement.
The model, called E(3)-Pose, tracks the position of the fetal head while the scan runs so the scanner can adjust and the slices form a complete 3D image. Its distinguishing feature is that it does not try to learn the geometry of a head purely from data. The left-right symmetry of the fetal head is written into the network's architecture, which makes it easier for the model to work out which way the brain is facing. The work, led by MIT PhD student Ramya Muthukrishnan in Polina Golland's group, was presented at the European Conference on Computer Vision in 2026.
The project also ran on real hardware, which is where many medical imaging models stop. Muthukrishnan describes crashing the scanner many times while getting the model to run on it, with help from Borjan Gagoski at Boston Children's, who designed an extra MRI sequence to gather training data, and Aryn Lee, who recruited pregnant volunteers. The clinical motivation is concrete: Muthukrishnan recalls watching a pregnant patient abandon a scan halfway through after an hour in the machine, leaving the radiologist without a usable image. Ellen Grant of Boston Children's says the approach should lead to faster, more comfortable scans and better images.
The design lesson reaches beyond prenatal care. Building known structure, such as symmetry, into a model rather than hoping it learns it is a sound way to get reliable behaviour from limited medical data, and the authors suggest uses in other hard-to-image anatomy and even in tracking 3D objects with noisy sensors.
What is missing is evidence of clinical benefit. The MIT article reports no accuracy figures, no comparison with existing motion-correction methods and no count of patients scanned. The model was trained mainly on third-trimester fetuses, while the earlier second trimester, when fetuses are smaller and move more, is where detection matters most and where the team says it wants to go next. For now this is a promising method with a working prototype, not a measured improvement in diagnosis.