Prenatal MRI is one of medicine’s most valuable diagnostic tools, and one of its most frustrating. When a routine ultrasound flags a potential brain abnormality in a developing fetus, an MRI is the logical next step. The problem is that fetuses move. And when they do, the scan falls apart.
A research team from MIT, Boston Children’s Hospital, and Inria has developed a deep learning model called E(3)-Pose that addresses this problem directly, not by making MRI faster or cheaper in the abstract, but by solving a specific, concrete failure mode that has long undermined diagnostic accuracy in fetal imaging.
The Physics of a Moving Target
To understand why fetal MRI is so difficult, it helps to understand how the imaging works. An MRI of the fetal brain is not a single photograph. It is a stack of two-dimensional image slices that, when assembled, form a three-dimensional volumetric picture of the brain. The quality of that picture depends entirely on each slice being taken from a consistent, predictable angle.
When the fetus moves, that consistency breaks down. Each slice may be captured from a different orientation, leaving gaps in coverage and misaligned angles in the final image. As first author Ramya Muthukrishnan, a PhD student at MIT, explains it: the slices end up “all over the place.” The radiologist reviewing the scan is then working with an incomplete picture, literally. Fine structural details that would indicate abnormal brain development become invisible or ambiguous. Missed diagnoses and incorrect diagnoses follow.
This is what most coverage of AI in medicine glosses over. The challenge is not that AI needs to be smarter in some general sense. It is that a very specific physical problem, fetal motion during image acquisition, has been degrading diagnostic reliability for years, and the solution requires understanding the physics of that problem, not just training a model on large datasets.
Symmetry as a Design Choice
What distinguishes E(3)-Pose from a conventional deep learning approach is where the physics lives in the model. Rather than feeding the model data and hoping it learns physical principles on its own, the researchers embedded those principles directly into the neural network’s architecture. Specifically, the left-right symmetry of the fetal head is built into the model’s structure. This makes it significantly easier for the system to determine the brain’s orientation accurately, even when the fetus is in motion.
This design philosophy, encoding known physical constraints into the model rather than leaving them to be inferred from data, is what Muthukrishnan’s supervisor, Professor Polina Golland, describes as combining “geometric concepts like equivariance and symmetry with the power of deep learning.” The result is a model that tracks fetal head position more reliably because it is not guessing at symmetry from examples. It already knows it.
The development process was not straightforward. Muthukrishnan came to the project with a background in computer science, not MRI physics. Implementing the model on actual scanner hardware, as opposed to running it on data in a computing environment, proved to be a different challenge entirely. She repeatedly crashed the scanner during testing. Co-author Borjan Gagoski, an assistant professor of radiology at Harvard Medical School and researcher at Boston Children’s Hospital, helped troubleshoot those failures and developed additional MRI acquisition sequences to support the model’s training. Research assistant Aryn Lee at Boston Children’s recruited the pregnant volunteers and managed scan acquisition.
The clinical motivation behind the work is direct. Muthukrishnan recalls observing a fetal MRI session at Boston Children’s Hospital early in her PhD, where a pregnant patient spent an hour in the scanner only to stop the scan halfway through due to discomfort, leaving the radiologist without a usable image. That experience, she has said, helped define the direction of her research. The frustration was shared, and the gap between what the technology could theoretically provide and what it actually delivered in practice was visible.
Why This Matters Beyond the Clinic
The implications of E(3)-Pose extend in two directions. The first is clinical and immediate. Ellen Grant, who directs the Fetal Neonatal Neuroimaging and Developmental Science Center at Boston Children’s Hospital, has noted that the work could automate fetal brain MRI, replacing a slow and manual process with faster, more comfortable scans and higher-quality images for radiologists. The current model is trained primarily on third-trimester fetuses. Muthukrishnan’s stated next goal is to extend it to second-trimester imaging, where the stakes are higher: younger fetuses move more, are smaller, and are harder to image, but earlier detection of brain abnormalities creates more options for intervention.
The second direction is broader. The same architectural approach, physics-informed neural networks that encode known geometric constraints, has potential applications in imaging other difficult anatomy and in fields like autonomous driving, where sensors must track three-dimensional objects through noisy, incomplete data.
Here is what that signals about AI development more generally: the most durable advances often come not from scaling models up, but from building domain knowledge into their foundations. A model that understands symmetry because symmetry is part of its structure will outperform one that has to learn it from examples, especially when the examples are limited, as they inevitably are in specialized medical imaging.
In Short
E(3)-Pose is a deep learning model designed to track fetal head movement during MRI scans, enabling clearer, more complete images of the developing brain. Its key innovation is architectural: physical principles, specifically the left-right symmetry of the fetal head, are embedded directly into the model rather than learned from data. Developed by researchers at MIT, Boston Children’s Hospital, and Inria, the system addresses a concrete diagnostic problem that has long caused missed and incorrect prenatal diagnoses. The broader lesson is that encoding what is already known about the physical world into AI systems can produce more reliable tools than training alone.
Based on reporting from MIT News AI.