Researchers at MIT, Boston Children’s Hospital and the French research institute Inria have developed a deep learning model called E(3)-Pose that tracks the position of a fetus’s head while it moves inside an MRI scanner. The goal, according to MIT News, is to let the scan adjust on the fly so that radiologists get a complete, high-resolution picture of the developing brain, and so that families are potentially spared repeat visits. The work was presented in a paper at the 2026 European Conference on Computer Vision (ECCV), a peer-reviewed venue; MIT’s release does not describe a clinical trial, and the model has been tested on volunteers.
The problem: a moving target
Prenatal care often starts with a routine ultrasound. When a sonographer flags a possible brain abnormality, doctors may order a fetal MRI for a closer look. But MRI is slow, and a fetus does not hold still. An MRI scan collects a series of two-dimensional “slices” that are stacked into a three-dimensional image of the head. If the head shifts mid-scan, the slices end up at inconsistent angles. First author Ramya Muthukrishnan, a PhD student at MIT, put it this way: “the slices are all over the place.”
The consequences are practical. Gaps in coverage can hide critical regions of the brain, and fine details need a good volumetric picture to be seen at all. Imperfect imaging can lead to missed or incorrect diagnoses. Muthukrishnan recalled shadowing a radiologist at Boston Children’s Hospital and seeing both sides frustrated: a pregnant mother uncomfortable after an hour in the scanner, who sometimes quit halfway, and a radiologist unable to obtain a complete, high-quality scan. That experience, she said, shaped her PhD.
The idea: build symmetry into the network
E(3)-Pose estimates the orientation, or pose, of the fetal head so the scanner can compensate. What the MIT team emphasizes is the model’s use of physics and geometry. Instead of leaving a network to discover regularities from data alone, the researchers embedded the left-right symmetry of the fetal head in the model architecture. That makes it easier for the network to work out the brain’s orientation accurately.
The “E(3)” in the name refers to the group of rigid motions in three-dimensional space, rotations, translations and reflections. Models designed to respect such symmetries are called equivariant: if the input is rotated or shifted, the output transforms in a matching way rather than needing to relearn the pattern for every position. In general, this can reduce how much training data a model needs and make its predictions more consistent. For an introduction to the family of image models this work builds on, see our explainer on the vision transformer.
From a trained model to a working scanner
A notable detail is that this research reached actual hardware. Muthukrishnan said she was surprised when she had to implement the model on the scanner, and recalled how many times she crashed it. Running software on scanner hardware, she said, “is a different beast” than training a model on a GPU.
She had help from clinical collaborators. Co-author Borjan Gagoski, an assistant professor of radiology at Harvard Medical School and a researcher at Boston Children’s, developed an MRI sequence that acquires additional scans to help train E(3)-Pose and helped troubleshoot the scanner. Co-author Aryn Lee, a research assistant at Boston Children’s, recruited the pregnant volunteers and helped acquire the scans. Senior authors include Elfar Adalsteinsson, Ellen Grant and Polina Golland. Benjamin Billot, a researcher at Inria and former postdoc in Golland’s group, helped with the initial method.
What the researchers say it could do
Golland said the work “elegantly combines geometric concepts like equivariance and symmetry with the power of deep learning to change how MRI scanners acquire the images,” with clear clinical impact. Grant, who directs the Fetal Neonatal Neuroimaging and Developmental Science Center at Boston Children’s, said it “will automate fetal brain MRI, replacing today’s slow, manual process with faster and more comfortable scans for mothers, while also providing higher-quality images for radiologists.” These are the researchers’ own expectations, not measured outcomes, and the release gives no data on diagnostic accuracy or scan time.
The team suggests wider uses too. E(3)-Pose could help image other hard-to-image anatomy, and even assist autonomous driving, where limited sensors must capture noisy 3D data to track objects.
Limits
MIT News describes one clear limitation: the model is trained primarily on third-trimester fetuses. Second-trimester and younger fetuses are small and move more because there is more space in the uterus, yet they matter more because earlier detection of brain abnormalities makes intervention easier. Muthukrishnan hopes to extend the model to those cases. Regulatory approval, testing across scanner models and hospitals, and evidence that it improves diagnoses would all be needed before routine use. Our overview of AI in healthcare risks and regulation outlines that path.
The work was supported by the MIT Jameel Clinic, the MIT Takeda Program, the Chou Family Transformative Research Fund, the National Institute of Biomedical Imaging and Bioengineering and the Eunice Kennedy Shriver National Institute of Child Health and Human Development.
What to watch next
- Results for second-trimester fetuses, which the team says it hopes to address.
- Published data on image quality and scan time compared with current methods.
- Testing at more hospitals and on more scanner types.
- Application of the same equivariant approach to other organs or sensors.
Key takeaways
- E(3)-Pose tracks the fetal head’s orientation during MRI so scans can adapt to motion.
- Left-right symmetry of the head is built into the model’s architecture.
- The paper was presented at ECCV 2026; the model has been implemented on a scanner with pregnant volunteers.
- It is trained mainly on third-trimester fetuses; earlier pregnancies are a goal.
- Claims about faster, more comfortable scans are the researchers’ expectations, not yet reported results.
Frequently asked questions
What does “equivariant” mean?
An equivariant model transforms its output predictably when its input is rotated, shifted or reflected, so it does not have to relearn each position.
Is E(3)-Pose used in hospitals today?
MIT News describes research tests with volunteers, not routine clinical use.
Why is fetal MRI hard?
The fetus moves during the scan, so the slices that form the 3D image can be misaligned and leave gaps.
