AI Enables Faster, More Precise Image Registration for Medical Image Analysis
In This Article
- Image registration enables comparison of medical images by ensuring that the images occupy the same space in a coordinate system
- Researchers at Massachusetts General Hospital recently introduced an AI-based approach to image registration that is faster and more precise than was previously possible
- The new tool could benefit a wide range of research and even clinical applications
Imaging studies are critical in identifying differences in individuals' brains across populations or over time in a single person. Oftentimes, this would not be possible without a process called image registration, which facilitates the comparison of images by aligning them so they occupy the same space in a coordinate system.
Subscribe to the latest updates from Radiology Advances in Motion
Now, a team of researchers with the Athinoula A. Martinos Center for Biomedical Imaging at Massachusetts General Hospital has developed a new approach to image registration, dubbed SynthMorph, that takes advantage of artificial intelligence (AI) to achieve registration more robust, accurate, and faster than existing methods. They describe SynthMorph in a paper published in Imaging Neuroscience.
In the Q&A below, Malte Hoffmann, PhD, a researcher in the Laboratories for Computational Imaging at the Martinos Center and lead author of the study, discusses the new tool and its potential for a host of applications.
The study's other authors are Andrew Hoopes, Douglas N. Greve, PhD, Bruce Fischl, PhD, and Adrian V. Dalca, PhD.
Q. Why is image registration important for brain imaging?
Hoffmann: Image registration is a key technique in medical image analysis that estimates the spatial transformation from the anatomy in one image to the anatomy in another. Its fundamental goal is to aid the comparison of images, either from the same individual across time or between different individuals. For example, a common use case of registration is to remove differences due to head positioning between three-dimensional magnetic resonance imaging (MRI) brain scans acquired at different times so that any anatomical change is easier to see.
Q. What is SynthMorph? What advantages does it offer over conventional approaches to image registration?
Hoffmann: SynthMorph is an end-to-end brain registration tool that uses artificial intelligence to overcome some of the drawbacks of existing methods. To register specific anatomy, most methods rely on preprocessing techniques that remove irrelevant and often distracting image content. As structures like the neck move or deform independently of the brain, they can impinge on accuracy. In contrast, SynthMorph learns the anatomy brain registration seeks to align, enabling high accuracy and robustness without preprocessing.
One way SynthMorph differs from other AI methods is that we train it solely with synthetic images that include but also exceed the realistic range. This makes our tool largely applicable to image types unseen at training, whereas other AI methods tend to struggle with data that differs from the training distribution: for example, images acquired with another type of MRI scan.
Q. What types of studies are using the tool?
Hoffmann: SynthMorph enables users to harness the power of modern deep learning without machine-learning expertise or high-end computational resources. While we distribute a standalone registration tool on this website, we are also integrating it with FreeSurfer, a software suite for the analysis of brain MRI developed by our group. FreeSurfer outputs are part of the official data releases of large studies like the Alzheimer's Disease Neuroimaging Initiative (ADNI) or the Human Connectome Project (HCP) that include thousands of brain scans, and SynthMorph has the potential to reduce processing times substantially.
At Mass General, we are working with researchers in Radiation Oncology to integrate SynthMorph in a study evaluating the effect of dose rates on clinical outcomes of external beam radiation therapy.
Q. Do you have plans to implement SynthMorph clinically?
Hoffman: We plan to implement SynthMorph with longitudinal techniques that derive morphological metrics from structural MRI scans. These metrics include patterns of cortical atrophy that are indicative of Alzheimer's Disease years before onset. We hope to improve their sensitivity and accessibility with SynthMorph, to facilitate the clinical diagnosis of Alzheimer's Disease in hospitals with limited access to advanced biomarkers.
Learn more about the Laboratories for Computational Neuroimaging
Learn more about the Athinoula A. Martinos Center for Biomedical Imaging