AI Tool Improves Fetal MRI for Brain Abnormality Diagnosis
Posted on 22 Oct 2026
Fetal brain abnormalities can be difficult to diagnose before birth when fetal motion reduces imaging quality. Routine ultrasound may flag a possible abnormality, prompting follow-up fetal MRI, but head movement during acquisition can obscure critical regions of the developing brain. To improve visualization in these cases, researchers have developed an artificial intelligence model designed to make fetal MRI more reliable.
Developed by researchers at MIT, Boston Children’s Hospital, and Inria, the deep learning model E(3)-Pose helps produce high-resolution images of the fetal head despite movement during scanning. It addresses a common challenge in fetal MRI: aligning multiple two-dimensional image slices to reconstruct a three-dimensional view. By improving this alignment, the approach aims to give radiologists clearer images to support earlier and more accurate diagnosis of developmental anomalies.
E(3)-Pose incorporates physical principles directly into its neural network, embedding the fetal head’s left-right symmetry into the model’s architecture. This design helps the system determine the brain’s orientation as the fetus moves, supplementing what it learns from training data. More reliable orientation estimates are intended to improve alignment of slices acquired at unusual angles and reduce gaps in the reconstructed image.
The work was presented at the 2026 European Conference on Computer Vision. A paper on the method, titled “Equivariant symmetry-aware head pose estimation for fetal MRI,” is available on the arXiv preprint server. The project involved MIT, Boston Children’s Hospital, Harvard Medical School, and Inria. The model is currently focused mainly on third-trimester fetuses, and further development aims to improve imaging of second-trimester or younger fetuses.
The same motion-tracking approach could also be relevant to other anatomy that is difficult to image. The researchers also noted possible applications outside medicine, including systems that must track three-dimensional objects from limited and noisy sensor data.
Related Links
MIT
Boston Children's