The human fetus’s development is still a complicated process that is susceptible to genetic, environmental, and metabolic disturbances despite millions of years of evolution. Developmental problems could arise from these disruptions, beginning with the brain, which is one of the first organs to develop in the fetus and continues to develop after birth.
The sonographer will identify potential brain abnormalities for further magnetic resonance imaging (MRI) during a regular prenatal ultrasound. However, poor imaging might lead to missed or inaccurate diagnoses.
Boston Children’s Hospital, and Inria have created and trained a deep learning model called E(3)-Pose that helps image the fetus’s head while it is moving to produce a high-resolution image that the radiologist can review in order to aid in the early detection and treatment of these abnormalities. The work was presented in a paper at the European Conference on Computer Vision in 2026.
According to primary author Ramya Muthukrishnan, “one of the problems with having the baby move is that the slices are all over the place.” Muthukrishnan is a PhD candidate studying under Sunlin (1966) and Priscilla Chou Professor of Electrical Engineering and Computer Science Polina Golland, a principal investigator at the Computer Science and Artificial Intelligence Laboratory (CSAIL) and Abdul Latif Jameel Clinic for Machine Learning in Health (Jameel Clinic).
Muthukrishnan refers to a set of 2D pictures obtained during a fetal MRI scan as “slices.” The fetus’s head appears as a single, high-resolution, three-dimensional image when these two-dimensional planes are placed together. However, each 2D picture “slice” may be taken from a different angle if the head moves during imaging, which could lead to gaps in the final scan and an odd slice orientation. These issues could lead the doctor examining the MRI to overlook or misidentify a prenatal brain disease.
According to Muthukrishnan, “you need a very good volumetric picture of the brain to see very fine details.” “You need to be able to see well to detect abnormal development, and you can’t see critical regions if you have coverage gaps.”
E(3)-Pose’s emphasis on physics is what makes it especially effective in tracking fetal mobility. The neural network’s architecture incorporates the left-right symmetry of the fetal head, which makes it easier for the model to correctly identify the direction of the brain. This is in contrast to merely training the model to learn physics from the data.
Muthukrishnan had no prior experience with physics, much less MRI physics, when she started working on the project at the beginning of her PhD. She worked in computer science for the most part. “When I had to go and implement this on the scanner, I was surprised,” she remembers. “Training a model to run on data on a GPU is a different beast than putting things on hardware and getting it to work on hardware, so I was not prepared for how many times I would crash the scanner.”
Muthukrishnan persevered in spite of the difficulties. At the beginning of her PhD, she recalls shadowing a radiologist who performed fetal MRI scans at Boston Children’s Hospital and witnessing the shared frustration of both the patient and the radiologist: the radiologist was unable to obtain a complete, high-quality MRI to track the fetus’s development, and the pregnant mother felt uncomfortable after spending an hour in the scan, only to stop it midway.
“I felt like this was a pain point, so seeing frustration about it helped me decide what I wanted to do with my PhD,” Muthukrishnan adds.
Luckily, she had assistance: co-author Borjan Gagoski, a researcher at Boston Children’s Hospital and assistant professor of radiology at Harvard Medical School, created a sequence that gathers extra MRI scans to help train E(3)-Pose and collaborated with Muthukrishnan to troubleshoot the scanner whenever it crashed. Aryn Lee, a co-author and research assistant at Boston Children’s, helped get each MRI image and recruited the pregnant volunteers. Elfar Adalsteinsson, an electrical engineering and computer science professor; Ellen Grant, a radiology professor at Harvard Medical School and the Käthe Beutler Professor of Pediatrics; and Golland are senior authors who contributed to the project’s direction. Benjamin Billot, a researcher at Inria and a former postdoc in Golland’s lab, assisted Muthukrishnan in launching the first technique he created.
In order to alter how MRI scanners obtain images and ultimately assist doctors in identifying and treating issues early in development, Ramya’s work skillfully blends geometric ideas like equivariance and symmetry with the power of deep learning. According to Golland, “it’s an amazing research project with a clear clinical impact that is deep and fundamental in how it models image acquisition.”
Muthukrishnan’s 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,” says Grant, who oversees the Fetal Neonatal Neuroimaging and Developmental Science Center (FNNDSC) at Boston Children’s.
Applications for E(3)-Pose may go beyond motion tracking in fetal MRIs. It could help with imaging other difficult-to-image anatomy and even autonomous driving applications, where tracking 3D objects requires limited sensors to record noisy 3D data.
Since the model is mostly trained on third-trimester babies, Muthukrishnan’s current goal is to enhance E(3)-Pose so that it may assist with imaging second-trimester or younger pregnancies. According to Muthukrishnan, second-trimester fetuses are “so small, they move around a lot more because there’s more space in the uterus.” “They are more crucial to concentrate on because it is easier to intervene when brain abnormalities are identified early.”
The National Institute of Biomedical Imaging and Bioengineering, the Eunice Kennedy Shriver National Institute of Child Health and Human Development, the Jameel Clinic, the Takeda Program, and the Chou Family Transformative Research Fund all provided support for this work.