Predicting Brain Age from T2-FLAIR Captures White Matter Aging Associated with Cardiovascular Risks

1195 

Abstract Submission 

Cong Zang¹, Elizabeth Haddad², Gilsoon Park¹, Sook-Lei Liew³, Neda Jahanshad⁴, Hosung Kim⁵

¹University of Southern California, LOS ANGELES, CA, ²USC, Marina Del Rey, CA, ³University of Southern California, San Diego, CA, ⁴Imaging Genetics Center, Keck School of Medicine of University of Southern California, Los Angeles, California, ⁵University of Southern California, Los Angeles, CA

Cong Zang  
University of Southern California
LOS ANGELES, CA

Elizabeth Haddad  
USC
Marina Del Rey, CA

Gilsoon Park  
University of Southern California
LOS ANGELES, CA

Sook-Lei Liew, PhD, OTR/L  
University of Southern California
San Diego, CA

Neda Jahanshad, PhD  
Imaging Genetics Center, Keck School of Medicine of University of Southern California
Los Angeles, California

Hosung Kim  
University of Southern California
Los Angeles, CA

In brain age prediction (BAP) studies, machine learning, especially deep learning, is commonly used to estimate 'brain age' (BA). The brain age gap (BAG) is measured as the difference between predicted brain age and chronological age (CA) and offers a quantitative measure for assessing normal versus abnormal aging.
Various modalities of brain MRI data, including T1, T2-FLAIR, functional, and diffusion MRI, provide distinct features of brain structure and function that change with aging. T1 MRI is optimal for evaluating morphological changes in the gray matter (GM) and white matter (WM), while the visibility of lesions, such as white matter hyperintensities (WMH) or ischemic stroke lesions, is more pronounced on T2-FLAIR [8]. Recently, studies have employed T2-FLAIR to predict brain age trajectories from the whole WM volume [7] or WMH volumes [2]. However, they do not examine the relationship between the spatial distribution of WMH and aging. Given the distinctive association of cardiovascular diseases with WMHs in deep WM and periventricular WMHs [4], we aimed to examine brain age as predicted by the spatial distribution of WMH. We modeled medial surfaces generated at various depths from the WM-GM boundary to the ventricles and projected T2-FLAIR intensity values onto these medial surfaces. These values at different depth surfaces were then inputted into graph convolutional networks (GCN) to predict brain age. We hypothesize that the BAGs derived from T2-FLAIR signals sampled at various depth levels within WM represent WM-specific brain aging and will be associated with cardiovascular risks.

The UK BioBank dataset was used in this study (Fig. 1A). T1 images were processed through the CIVET pipeline (Fig. 1B) to construct inner cortical surfaces (namely WM surfaces) [9]. SynthSeg was used to run whole-brain segmentation [1]. We generated the Laplacian field vectors (LFV) from the WM surface to the ventricular boundary [5,6]. Vertices on the GM-WM boundary were deformed along the LFV, and medial surfaces were constructed using vertices at the given depth ratios (Fig. 1C) whose mesh topology was kept as the original GM-WM boundary (SurfStat, 2009). T2-FLAIR images were sampled onto the nearest neighboring vertices across all medial surfaces to form a feature matrix (5124 vertices x 9 surfaces, Fig. 1D).

We then randomly split the dataset into i) training data (7631 subjects, age = 62.7±7.4 years, range 46-80) and test data (3370 subjects, age = 62.1±7.1 years, range = 48-77). The feature matrix was normalized to (0,1) before feeding it into the GCN. In this preliminary study, one BAG representing the whole WM was calculated. However, with this method, we could compute BAGs for different WM depth surfaces and left and right hemispheres separately. We then applied a general linear model, including risk factors, diseases and life habits as covariates. Results are Bonferroni-corrected by multiplying the original p-value with the number of comparisons (23 in total) made.

Our WM-BAP model predicted brain age with a mean absolute error of 3.08 on test data. Various cardiovascular risk factors such as hypertension, high BMI, high WMH burden, high blood pressure, and high heart rate were associated with increased WM-BAG (Fig 2).

Our BAP model based on T2-FLAIR signals at various WM depths captured accelerated aging associated with cardiovascular risk factors and diseases. WMH at different depths may impact brain aging differently, e.g. periventricular WMH may be associated with mild parenchymal changes, possibly due to nonischemic causes such as ependymal gliosis, demyelination, and discontinuation of the subependymal lining, while deep WM WMH may show more severe parenchymal changes, considered to be of ischemic origin [3, 10]. This study establishes a foundation for WM BAP studies, which could enable future analyses of aging trajectories in specific depths of WM and their contribution to cerebrovascular diseases.

Aging ¹

Classification and Predictive Modeling

White Matter Anatomy, Fiber Pathways and Connectivity ²

Anatomical MRI

Multi-Modal Imaging

Aging

Cerebrovascular Disease

Data analysis

Machine Learning

MRI

STRUCTURAL MRI

White Matter