Despite millions of years of evolution, human fetal development remains a highly complex process that is vulnerable to genetic, environmental, and metabolic disturbances, often leading to developmental abnormalities. The brain is one of the first organs to form in the fetus and continues to develop after birth. In routine prenatal ultrasound, possible brain anomalies are flagged for follow‑up magnetic resonance imaging (MRI), but imperfect imaging can cause missed or incorrect diagnoses.
To address this, researchers from MIT, Boston Children’s Hospital, and Inria have built a deep‑learning model called E(3)-Pose that captures high‑resolution images of the fetal head while it moves. The key innovation is embedding physical symmetry directly into the network architecture: left‑right symmetry of the fetal head is encoded as an equivariant constraint, enabling the model to infer the correct orientation even under motion blur.
During a fetal MRI scan, a series of 2‑D slices are acquired and stacked to form a 3‑D volume. If the head shifts during acquisition, each slice may be taken from a different angle, creating gaps and irregular slice orientation in the final volume, which hampers a radiologist’s ability to see fine‑grained structures.
E(3)-Pose excels because:
- It hard‑codes left‑right symmetry, allowing rapid pose recovery despite motion;
- Equivariant convolutions capture spatial transformations, reducing reliance on massive labeled datasets;
- The model is tightly coupled with the MRI pulse sequence, enabling real‑time pose correction on the scanner.
First author Ramya Muthukrishnan entered the project with a computer‑science background and no prior physics or MRI experience. Implementing the model on the scanner caused numerous crashes, teaching her that “running code on hardware is a different beast than training on a GPU.” Observing pregnant mothers endure long, uncomfortable scans that often end prematurely, and radiologists struggling to obtain complete high‑quality images, motivated her to focus on improving scan efficiency and image fidelity.
Co‑author Borjan Gagoski (Harvard Medical School, radiology) designed an additional MRI sequence to generate training data for E(3)-Pose and helped troubleshoot scanner crashes. Research assistant Aryn Lee recruited pregnant volunteers and performed the scans. Senior collaborators include MIT EECS professor Elfar Adalsteinsson, Harvard pediatric‑radiology professor Ellen Grant, and MIT CSAIL principal investigator Polina Golland. Benjamin Billot from Inria contributed the initial equivariant geometry framework.
Golland praised the work: “Ramya elegantly combines geometric concepts such as equivariance and symmetry with deep learning to change how MRI scanners acquire images, with clear clinical impact.” Grant added that E(3)-Pose will automate fetal brain MRI, replacing today’s slow, manual process with faster, more comfortable scans for mothers while delivering higher‑quality images for radiologists.
Beyond fetal MRI, E(3)-Pose could aid imaging of other hard‑to‑capture anatomy and even autonomous‑driving scenarios where limited sensors must reconstruct noisy 3‑D data. Currently trained on third‑trimester fetuses, the team aims to extend the model to second‑trimester or earlier stages, as earlier detection of brain abnormalities enables easier intervention.
The work is supported by the MIT Jameel Clinic, MIT Takeda Program, Chou Family Transformative Research Fund, the National Institute of Biomedical Imaging and Bioengineering (NIBIB), and the Eunice Kennedy Shriver National Institute of Child Health and Human Development.
Review: E(3)-Pose demonstrates how embedding symmetry priors into neural networks can robustly handle fetal motion, offering a powerful example of geometry‑aware deep learning that promises more reliable prenatal diagnostics.