Acquisition-independent machine learning for morphometric analysis of underrepresented aging populations with clinical and low-field brain MRI - Project Summary
Title:
Acquisition-independent machine learning for morphometric analysis of underrepresented aging populations
with clinical and low-field brain MRI
Summary:
Magnetic resonance imaging (MRI) has revolutionized research of the human brain, by providing a window to
the living brain in healthy aging and disease. A key aspect of this revolution has been the ability to obtain
precise morphometric measurements from brain MRI using software packages like FreeSurfer (developed by
our lab), FSL, AFNI, or SPM. These packages rely on computer algorithms that work best with isotropic data,
1 mm MP-RAGE scans. Unfortunately, clinical scans are generally highly anisotropic (e.g., 6 mm spacing
between slices), precluding automatic morphometric analysis with the aforementioned packages. Images
acquired with portable and non-portable low-field scanners also suffer from the same limitation.
The inability to process clinical MRI prevents the extraction of precise morphometric measurements from
MRI studies in the clinic (quantitative imaging), as well as from low-field scans that may be the only imaging
alternative in medically underserved regions, e.g., rural areas or developing countries. Crucially, this inability
also precludes the analysis of millions of scans that are sitting in the PACS of hospitals around the world,
including large amounts of images and associated clinical metadata from populations that are typically
underrepresented in neuroimaging studies (e.g., Black, Hispanic), thus hindering progress in aging research.
In this project, we propose to develop AI methods that can turn clinical or low-field MRI into isotropic scans
of reference contrast (a 1 mm MP-RAGE). Importantly, the methods will: (i) be adaptive to the number of input
MR sequences, as well as their orientation, contrast and resolution; (ii) be robust against aging-related
pathology (atrophy, white matter lesions); and (iii) not require retraining. These features will enable application
any MR dataset without specialized hardware or machine learning expertise. The resulting synthetic scans can
be used for a wide array of existing morphometrics analyses, e.g., segmentation, volumetry, registration,
longitudinal analysis, cortical thickness and parcellation, and many more. Another key feature of the framework
is the fact that it yields harmonized images, which reduces bias across sites and pulse sequences.
We will validate the tools with: (i) synthetically downsampled scans from a number of public datasets
covering a diverse population; and (ii) a dedicated, diverse, comprehensive dataset of multi-modal MRI,
comprising paired research, clinical and low-field scans, acquired specifically for this project. We will carefully
assess the biases in our developed tools and try to mitigate them. We will apply the final version of the tools to
a large-scale study of a clinical aging cohort from Massachusetts General Hospital, as well as to two clinical
studies with portable MRI. Both the tools and the new dataset will be made publicly through FreeSurfer (60,000
worldwide licenses), thus enabling researchers worldwide to analyze large clinical datasets with sample sizes
much higher than those achieved in current research studies. Therefore, our tools promise to increase our
understanding of the human brain in normal aging and in disease, particularly in underrepresented populations.
