AI model that embeds physics of symmetry improves fetal brain MRI, researchers report

E(3)-Pose tracks fetal head motion in real time to produce clearer 3D images for prenatal diagnosis

By LineZotpaper

Published20 October 2026

Read Time3 min

Researchers at MIT, Boston Children's Hospital, and Inria have developed a deep learning model called E(3)-Pose that uses physics-based symmetry to track fetal head movement during MRI scans, producing higher-quality images that could help clinicians detect brain abnormalities earlier. The work, presented at the 2026 European Conference on Computer Vision, addresses a longstanding challenge in fetal imaging: even slight motion by the baby can produce gaps or distortions in the final 3D scan, potentially leading to missed or incorrect diagnoses.

Fetal MRI is a critical tool for monitoring brain development, but it requires the fetus to remain still for extended periods. When the baby moves, the series of 2D image slices taken during the scan can become misaligned, creating coverage gaps that obscure fine anatomical details needed to detect abnormal development.

E(3)-Pose tackles this by incorporating the physical principle of equivariance — in this case, left-right symmetry of the fetal head — directly into the neural network architecture. This allows the model to more accurately estimate the brain's orientation as it shifts, even when the motion is complex.

"There are very fine details that you need a very good volumetric picture of the brain to see," said first author Ramya Muthukrishnan, an MIT PhD student in the lab of Professor Polina Golland. "If you have coverage gaps, you can't see critical regions, and you need to be able to see well to detect abnormal development."

The project began after Muthukrishnan shadowed a radiologist at Boston Children's Hospital and observed the frustration shared by pregnant patients and clinicians. Patients often had to spend an hour in the MRI scanner only to abort the scan midway, while radiologists were left without a complete, high-quality image.

Co-author Borjan Gagoski, an assistant professor of radiology at Harvard Medical School and researcher at Boston Children's, developed a supplemental MRI sequence to help train the model. Aryn Lee, a research assistant at Boston Children's, recruited pregnant volunteers and helped acquire the scans.

"Ramya's work elegantly combines geometric concepts like equivariance and symmetry with the power of deep learning to change how MRI scanners acquire the images," said senior author Polina Golland, who is a principal investigator at the MIT Abdul Latif Jameel Clinic for Machine Learning in Health and CSAIL. Ellen Grant, director of the Fetal Neonatal Neuroimaging and Developmental Science Center at Boston Children's, added that the work "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 to interpret."

E(3)-Pose is currently trained primarily on third-trimester fetuses. Muthukrishnan hopes to extend it to second-trimester scans, where the fetuses are smaller and move more, making them both harder to image and more important for early intervention.

The research 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.

Analysis

Why This Matters

Fetal brain abnormalities are often missed or misdiagnosed due to motion artifacts in MRI, causing delays in treatment or unnecessary follow-ups.
A more reliable imaging method could reduce scan time and discomfort for pregnant patients, while giving radiologists clearer data for diagnosis.
The same approach could be applied to other moving anatomy or even autonomous driving, where noisy 3D sensor data must track moving objects.

Background

Fetal MRI is commonly used to confirm or clarify findings from prenatal ultrasound, especially for brain development. However, the fetus can move unpredictably, and current motion-correction techniques often require manual intervention or post-processing. The E(3)-Pose model embeds geometric priors — specifically, the symmetry of the fetal head — into the neural network so that it can estimate pose from limited data without needing extensive training on motion examples. This approach is part of a broader trend in AI research where physical laws are built into model architectures rather than learned from data.

Key Perspectives

Researchers and clinicians: The team sees E(3)-Pose as a way to automate and improve fetal brain MRI, making scans faster and more comfortable while producing higher-quality images. They emphasize the clinical need, especially for early detection in the second trimester. Patients and pregnant mothers: The current process is described as frustrating, with long scan times that often end prematurely. A faster, more robust scan would reduce physical discomfort and anxiety. Critics and skeptics: The model is so far tested only on third-trimester fetuses. Second-trimester imaging, where the clinical need is greatest, presents greater challenges due to smaller size and more movement. Broader validation and integration into hospital workflows are still ahead.

What to Watch

Extension of E(3)-Pose to second-trimester fetuses, as Muthukrishnan plans.
Clinical trials or deployment at Boston Children's Hospital and other sites to test real-world reliability.
Adaptation of the equivariance-based approach to other MRI applications (e.g., imaging other moving organs) and to autonomous driving systems.

Sources

Why teaching AI about symmetry could make prenatal diagnosis more reliable · MIT News

Checked

12 claims in this story checked against 1 source: 11 supported.

Show each claim

SupportedAI model that embeds physics of symmetry improves fetal brain MRIMIT News: “left-right symmetry of the fetal head is embedded in the model architecture”
SupportedE(3)-Pose tracks fetal head motion in real time to produce clearer 3D images for prenatal diagnosisMIT News: “helps image the fetus’s head while it’s in motion to capture a high-resolution image”
Supported"There are very fine details that you need a very good volumetric picture of the brain to see," said first author Ramya Muthukrishnan, an MIT PhD student in the lab of Professor Polina Golland.MIT News: “There are very fine details that you need a very good volumetric picture of the brain to see”
Supported"If you have coverage gaps, you can't see critical regions, and you need to be able to see well to detect abnormal development."MIT News: “If you have coverage gaps, you can’t see critical regions, and you need to be able to see well to detect abnormal development.”
SupportedThe project began after Muthukrishnan shadowed a radiologist at Boston Children's Hospital and observed the frustration shared by pregnant patients and clinicians.MIT News: “She remembers shadowing a radiologist who performed fetal MRI scans at Boston Children’s Hospital at the start of her PhD and observing the shared frustration of the patient and radiologist”
Partly supportedPatients often had to spend an hour in the MRI scanner only to abort the scan midwayMIT News: “the pregnant mother felt uncomfortable after spending an hour in an MRI, only to quit the scan halfway”
SupportedCo-author Borjan Gagoski, an assistant professor of radiology at Harvard Medical School and researcher at Boston Children's, developed a supplemental MRI sequence to help train the model.MIT News: “Co-author Borjan Gagoski, an assistant professor of radiology at Harvard Medical School and a researcher at Boston Children's Hospital, developed a sequence that acquires additional MRI scans to assist with training E(3)-Pose”
SupportedAryn Lee, a research assistant at Boston Children's, recruited pregnant volunteers and helped acquire the scans.MIT News: “Co-author and Boston Children's Research Assistant Aryn Lee recruited the pregnant volunteers and assisted with acquiring each of the MRI scans.”
Supported"Ramya's work elegantly combins geometric concepts like equivariance and symmetry with the power of deep learning to change how MRI scanners acquire the images," said senior author Polina Golland.MIT News: “Ramya’s work elegantly combines geometric concepts like equivariance and symmetry with the power of deep learning to change how MRI scanners acquire the images”
SupportedEllen Grant added that the work "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 to interpret."MIT News: “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 to interpret.”
SupportedE(3)-Pose is currently trained primarily on third-trimester fetuses.MIT News: “the model is primarily trained on third trimester fetuses.”
SupportedMuthukrishnan hopes to extend it to second-trimester scans, where the fetuses are smaller and move moreMIT News: “Muthukrishnan hopes to improve E(3)-Pose so that it can help with imaging second-trimester or younger fetuses... Second-trimester fetuses are 'so small, they move around a lot more'”