Geometric symmetry tames fetal motion, a model ends MRI image disruption of newborn brains

Embryonic examinations inside MRI machines face a persistent clinical dilemma: the fetus’s repeated head movements within the womb blur three-dimensional image capture, and make the two-dimensional slice images spaced and angled inconsistently, leaving visual gaps that obscure critical neural details. This technical and physical challenge prompted researchers from the Massachusetts Institute of Technology, Boston Children’s Hospital and the Inria Institute to present a deep-learning model they called “E3BOS”, a system capable of tracking the fetal head motion in real time during the scan and reconstructing a high-resolution volumetric image for diagnostic radiology staff to examine.

The work was presented at the 2026 European Conference on Computer Vision, based on an innovative treatment of the motion problem. Rather than training the neural network to infer physical principles solely by processing massive data sets, the team incorporated physical laws and the concept of geometric equivalence directly into the network’s architecture, focusing on the natural symmetry between the right and left sides of the fetal head. This geometric grounding gave the model an immediate ability to determine the head’s orientation and adjust the slice trajectory without resorting to random guessing.

Embedding physical symmetry within the neural architecture turned image processing from statistical guessing into coherent geometric calibration.Lead researcher and PhD candidate at MIT, Ramya MuthokriShnan, explains that acquiring a complete three-dimensional image of the fetal brain requires stitching together dozens of two-dimensional slices with extreme precision. When the fetal head moves, each slice is captured at a completely different angle, creating coverage gaps that prevent physicians from inspecting the finest anatomical details and increase the likelihood of error in diagnosing early brain malformations that are expected to be detected accurately after conversion from routine ultrasound examinations.

Moving this model from a GPU-based training environment to practical deployment inside MRI scanners imposed a demanding real-world test on the team. MuthokriShnan recounts her experience accompanying radiologists at Boston Children’s Hospital, where pregnant mothers sometimes spent up to an hour inside the narrow scanner chamber in discomfort, and the scan often halted midway before the required coverage was completed. When attempting to run the model directly on the scanner hardware, the device failed repeatedly, as running the software on medical equipment demands timing and geometric constraints that exceed those of isolated servers. Assistant Professor of Radiology at Harvard Medical School, Boryan Gagowski, took charge of developing additional scan sequences and troubleshooting scanner faults on site, while researcher Arin Lee helped recruit the pregnant volunteers.

Automating fetal-motion tracking shortens examination time and provides mothers with a comfortable clinical experience that shields them from repeat sessions.Professor Ellen Grant, director of the Center for Neuroscience and Embryonic and Neonatal Development at Boston Children’s Hospital, confirms that this innovation replaces slow manual procedures with a high-precision automated scan. The researcher is currently, under the supervision of Professor Polina Golland at the Abdul Latif Jameel AI in Health Clinic, expanding the model’s capabilities, which was originally trained on third-trimester fetuses, to include second-trimester fetuses, a stage in which fetal movement increases as uterine space expands, yet also represents the most critical window for early medical intervention in addressing developmental malformations.

This technological shift carries direct operational and clinical implications for healthcare in the Gulf, Egypt and the region as a whole. Hospitals and maternal-child health centers that have invested heavily in acquiring the latest MRI scanners constantly encounter high repeat-scan rates and lost operating time of valuable equipment due to fetal motion and mothers’ inability to complete lengthy sessions. Turning the problem into a software solution that tracks motion at the level of scan sequences means that radiology departments and specialized centers can boost the efficiency of existing hardware and lower the cost of repeat imaging without needing to purchase new scanners. It also gives neurologists and diagnostic radiologists in the region greater capacity to make decisive decisions about complex fetal anomalies at early stages, turning computer-vision algorithms from isolated research into a clinical tool that supports the throughput of radiology departments and eases the burden on pregnant mothers.