Symmetry-aware AI model tracks fetal motion during MRI scans
MIT, Boston Children's Hospital and Inria researchers built E(3)-Pose, a deep learning model that adjusts fetal MRI scans in motion using head symmetry baked into its architecture.
The story in focus
Researchers at MIT, Boston Children's Hospital and Inria developed E(3)-Pose, a deep learning model that tracks fetal head motion during MRI by embedding left-right symmetry in its architecture, aiming to produce cleaner 3D images and reduce repeat scans. A paper was presented at the 2026 European Conference on Computer Vision.
- E(3)-Pose embeds the fetal head's left-right symmetry directly into the neural network architecture
- The model corrects slice misalignment when the fetus moves during 2D MRI acquisition
- Training focused mainly on third-trimester fetuses; younger fetuses are a stated next target
The imaging problem
Fetal MRI builds a high-resolution 3D image from stacked 2D slices. If the fetus's head shifts during scanning, slices can be captured at different angles, leaving coverage gaps and unusual slice orientation. MIT PhD student Ramya Muthukrishnan, quoted in an MIT News article, said the movement problem means slices are scattered, and missing coverage can hide critical brain regions a radiologist needs to assess.
Symmetry built into the model
E(3)-Pose differs from models trained to infer physics from data: the physics principle of left-right head symmetry is embedded in the network architecture, which the researchers say helps the model identify brain orientation accurately. The work was presented at the 2026 European Conference on Computer Vision. Co-author Borjan Gagoski developed an additional MRI acquisition sequence to support training, and Aryn Lee recruited pregnant volunteers and helped acquire scans.
Next milestone
The model has been trained mainly on third-trimester fetuses. Muthukrishnan said she hopes to improve E(3)-Pose for second-trimester or younger fetuses, which she described as smaller, more mobile and more important to image because earlier detection of brain abnormalities makes intervention easier. Ellen Grant of Boston Children's said the work would automate fetal brain MRI, replacing a slow manual process. The researchers also suggest possible use in other hard-to-image anatomy and autonomous driving.
What makes E(3)-Pose better at tracking fetal motion than data-only models?
Its architecture embeds the physics principle of left-right symmetry of the fetal head, which the researchers say helps it accurately identify brain orientation and correct misaligned slices.
Sources & further reading
Help keep the record accurate. Send a correction.
