Enhancing Fetal MRI Accuracy Using Symmetry-Aware Deep Learning
A new AI model, E(3)-Pose, improves prenatal brain imaging by correcting motion-induced artifacts.
1The Fetal Imaging Challenge
Fetal brain development is highly susceptible to genetic and environmental disruptions, necessitating precise prenatal monitoring. Standard MRI scans often suffer from motion artifacts because fetuses move during the procedure. These movements create inconsistent 2D slices, leading to gaps in volumetric data that can cause radiologists to miss or misdiagnose critical developmental pathologies.
2Introducing E(3)-Pose Architecture
Researchers from MIT, Boston Children’s Hospital, and Inria developed E(3)-Pose to address motion-related imaging failures. Unlike standard models that learn physics solely from data, this deep learning framework embeds the principle of left-right symmetry directly into its architecture. This geometric constraint allows the model to accurately track the fetal head's orientation.
3Bridging Physics and Hardware
Implementing the model on actual MRI hardware proved significantly more complex than training it on GPU-based data. Ramya Muthukrishnan, the lead author, frequently encountered system crashes while integrating the software with scanner hardware. Collaborating with radiologist Borjan Gagoski was essential to troubleshoot these technical hurdles and successfully acquire the necessary training data.
4Clinical Workflow Improvements
The current manual process for fetal MRI is slow and often uncomfortable for pregnant patients, frequently leading to incomplete scans. By automating the alignment of 2D slices, E(3)-Pose aims to shorten scan times and improve patient comfort. This efficiency ensures that radiologists receive high-quality, complete 3D images for more reliable clinical interpretations.
5Expertise and Collaborative Research
The project was guided by a multidisciplinary team including Polina Golland, Elfar Adalsteinsson, and Ellen Grant. Their collective expertise in electrical engineering, computer science, and pediatric radiology enabled the integration of equivariance concepts into the model. This collaboration bridged the gap between fundamental machine learning research and practical, high-impact clinical medical applications.
6Expanding Future Capabilities
While E(3)-Pose is currently optimized for third-trimester fetuses, researchers are working to adapt it for second-trimester imaging. Younger fetuses move more frequently due to increased uterine space, making them harder to capture. Improving detection in these earlier stages is a priority, as it allows for earlier medical intervention and better developmental outcomes.
7Broader Technological Implications
The potential applications for E(3)-Pose extend well beyond prenatal diagnostics. The model’s ability to handle noisy 3D data and track moving objects could benefit other medical imaging fields. Furthermore, the underlying technology holds promise for autonomous driving systems, where sensors must accurately interpret 3D environments despite constant motion and limited data inputs.
8Research Funding and Support
This research received support from several institutions, including the MIT Jameel Clinic and the MIT Takeda Program. Additional funding was provided by the Chou Family Transformative Research Fund and the National Institute of Biomedical Imaging and Bioengineering. These resources supported the development of a tool that fundamentally changes how MRI scanners acquire images.
Original article: MIT News — AI — Enhancing Fetal MRI Accuracy Using Symmetry-Aware Deep Learning