A new generation of “low-field” MRI systems is transforming the field of medical imaging. Relatively inexpensive, portable and unhampered by special power requirements, they promise greater accessibility to MRI in underserved areas and allow point-of-care imaging in cases where moving a patient can be dangerous to their health. But because of the weaker magnetic fields the systems use, the images they produce are often more difficult to interpret.
Enter Juan Eugenio Iglesias, PhD, an associate professor at the Martinos Center for Biomedical Imaging in the MGB Department of Radiology. Iglesias and his team are developing robust artificial intelligence (AI) tools that are helping overcome the obstacles associated with portable MRI, and thus facilitating translation of the technology into the clinic.
In a recent conversation, the Martinos researcher reflected on the transformative power of AI in medical imaging, and in low-field MRI, in particular, and shared his thoughts on what’s ahead. Here’s what we learned.
What development in AI and imaging feels most transformative right now — and why?
Well, it is hard to argue against progress in generative AI these days. But, if we are talking about medical imaging, I think that the most transformative development right now is the maturation of foundation models and self-supervised training. These models can learn robust representations from large volumes of unlabeled MRI data, which is especially important in low-field imaging where annotated datasets are scarce. Instead of training narrow models for single tasks, we’re beginning to see systems that generalize across denoising, reconstruction, segmentation, and pathology detection.
What need do you wish we could meet today, but the tools or methods aren’t quite there yet?
If you grant me two wishes, I would say: (a) that we could deliver fully reliable, automated triage-level interpretation at the bedside across pathologies like stroke or hemorrhage with performance matching high-field MRI; and (b) that we could measure subtle change longitudinally as accurately as we can with high-field scanners, which would be a game changer for clinical trials of diseases like Alzheimer’s.
What is the biggest obstacle to overcome in the near term, or the biggest bottleneck slowing progress?
I believe that data remains the biggest bottleneck. Low-field brain MRI datasets are not just “corrupted high-field MRI data” that we can easily and accurately simulate from high-field; they have their own contrast mechanisms and artifacts. We need more low-field data but, because portable low-field systems are newer, we do not yet have the decades of accumulated imaging data that fuel algorithm development.
Another key obstacle is integration into clinical workflow. Even if an AI model performs well technically, it must fit seamlessly into radiology and neurocritical care processes, including regulatory approval and clinician trust. It is not only about performance on our internal datasets, but also real-world validation and explainability.
What excites you most about where the field is heading over the next decade?
This is an easy one: democratization of neuroimaging. Portable low-field MRI, enhanced by AI, has the potential to move brain imaging beyond major healthcare and academic centers into community hospitals, rural settings, and global health environments where conventional MRI is impractical. We’re shifting from centralized, infrastructure-heavy imaging toward point-of-care intelligence. Over the next decade, I believe we will see AI-native imaging systems where acquisition parameters adapt in real time based on model feedback, scans are optimized for specific clinical questions, and diagnostic insights are generated at the bedside within minutes. That convergence of portable hardware, intelligent reconstruction, and clinical decision support could fundamentally reshape how and where we diagnose brain disease.
