Advanced diffusion modeling classifies FLAIR white matter hyperintensity types in an aging cohort
Poster No:
1602
Submission Type:
Abstract Submission
Authors:
Kelly Chang1, Luke Burke2, Nina LaPiana1, Bradley Howlett1, David Hunt1, Margaret Dezelar2, Jalal Andre1, James Ralston2, Ariel Rokem1, Christine Mac Donald1
Institutions:
1University of Washington, Seattle, WA, 2Kaiser Permanente Washington Health Research Institute, Seattle, WA
First Author:
Co-Author(s):
Introduction:
White matter hyperintensities (WMH) in fluid-attenuated inversion recovery (FLAIR) imaging are used as an indicator of clinical conditions ranging from multiple sclerosis to cerebrovascular disease (Ferris et al., 2022; Preziosa et al., 2023). However, the biophysics underlying FLAIR WMH is only partially understood, and current diagnostic standards rely on cumbersome visual inspection. In contrast, advanced diffusion MRI (dMRI) modeling, such as multi-shell and high angular resolution imaging, provides biophysically interpretable tissue properties. To study the relationships between biophysical tissue properties and FLAIR WMH in aging, we used a combination of unsupervised and supervised machine learning to classify the distribution of tissue properties within WMH regions of interest (ROI) in a sample from the Adult Changes in Thought (ACT) study (Kukull et al., 2002).
Methods:
143 participants (ages 70 - 103, mean age = 80.23; 80 females) underwent T1-weighted (T1w), T2-weighted (T2w), and FLAIR structural imaging as well as multi-shell diffusion MRI.
We used a neural network (Forooshani et al., 2022) to automatically segment WMH voxels from FLAIR images. WMH regions of interest (ROIs) were defined to be a cluster of contiguous WMH voxels. The WMH ROIs were then categorized as periventricular if the ROI was adjacent to the lateral ventricles – taking care to exclude voxels immediately adjacent (within 1 mm) to the ventricles to avoid partial volume effects – or as deep otherwise. The remaining white matter volume was categorized as normal-appearing white matter (NWM).
Diffusion modeling was performed on the QSIprep-0.18.1 (Cieslak et al., 2021) preprocessed data. Diffusional kurtosis imaging (DKI; Jensen et al., 2005), free-water diffusion tensor imaging (FWDTI; Hoy et al., 2014) and mean apparent propagator MRI (MAPMRI; Özarslan et al., 2013) models were fit using the Diffusion Imaging in Python (DIPY) software library. The neurite orientation dispersion and density imaging (NODDI) model (Zhang et al., 2012) was fit using the accelerated microstructure via convex optimization (AMICO) library. We derived several metrics from these models (and T1w/T2w), which were normalized to each participant's average within their white matter volume.
We used a neural network (Forooshani et al., 2022) to automatically segment WMH voxels from FLAIR images. WMH regions of interest (ROIs) were defined to be a cluster of contiguous WMH voxels. The WMH ROIs were then categorized as periventricular if the ROI was adjacent to the lateral ventricles – taking care to exclude voxels immediately adjacent (within 1 mm) to the ventricles to avoid partial volume effects – or as deep otherwise. The remaining white matter volume was categorized as normal-appearing white matter (NWM).
Diffusion modeling was performed on the QSIprep-0.18.1 (Cieslak et al., 2021) preprocessed data. Diffusional kurtosis imaging (DKI; Jensen et al., 2005), free-water diffusion tensor imaging (FWDTI; Hoy et al., 2014) and mean apparent propagator MRI (MAPMRI; Özarslan et al., 2013) models were fit using the Diffusion Imaging in Python (DIPY) software library. The neurite orientation dispersion and density imaging (NODDI) model (Zhang et al., 2012) was fit using the accelerated microstructure via convex optimization (AMICO) library. We derived several metrics from these models (and T1w/T2w), which were normalized to each participant's average within their white matter volume.
Results:
We found that periventricular WMH were characterized by greater mean diffusivity and extracellular water content, and less myelination (quantified as T1w/T2w ratio) than NWM (Figure 1A). Whereas deep WMH was more similar to NWM, except for a slightly greater mean diffusivity and a large decrease in myelination as indicated by the T1w/T2w ratio (Figure 1B).
We performed a principal components analysis (PCA) on the biophysical metrics, where the first two components explained 83.28% of the variance in the dataset (Figure 2A). A logistic regression classifier was trained on the first two PCs was able to classify periventricular and deep WMH ROIs with an average cross-validation accuracy of 77.84% (± 0.71%). We confirmed the classification performance by training a separate logistic regression classifier on the biophysical metrics, while controlling for age (refer to Figure 2B for model coefficients). The biophysical logistic regression had an average cross-validation classification accuracy of 80.64% (± 0.89%).
We performed a principal components analysis (PCA) on the biophysical metrics, where the first two components explained 83.28% of the variance in the dataset (Figure 2A). A logistic regression classifier was trained on the first two PCs was able to classify periventricular and deep WMH ROIs with an average cross-validation accuracy of 77.84% (± 0.71%). We confirmed the classification performance by training a separate logistic regression classifier on the biophysical metrics, while controlling for age (refer to Figure 2B for model coefficients). The biophysical logistic regression had an average cross-validation classification accuracy of 80.64% (± 0.89%).
Conclusions:
This study highlights the potential of advanced dMRI modeling in characterizing the biophysical properties of periventricular and deep WMH. The biophysical metric patterns indicate that periventricular and deep WMH tissue begins to resemble ventricles more than NWM, particularly in the case of periventricular WMH. Furthermore, the high classification accuracy achieved by the linear SVM emphasized the differences in underlying tissue properties of periventricular and deep WMH, suggesting pathological differences between the two WMH types.
Lifespan Development:
Aging 2
Modeling and Analysis Methods:
Diffusion MRI Modeling and Analysis 1
Keywords:
ADULTS
Aging
Data analysis
Modeling
MRI
STRUCTURAL MRI
White Matter
1|2Indicates the priority used for review
Provide references using author date format
Ferris, J. K. (2022). "In vivo myelin imaging and tissue microstructure in white matter hyperintensities and perilesional white matter", Brain communications, vol. 4, no. 3, fcac142
Forooshani, P. M. (2022). "Deep Bayesian networks for uncertainty estimation and adversarial resistance of white matter hyperintensity segmentation", Human Brain Mapping, vol. 43, no. 7, pp. 2089-2108
Hoy, A. R. (2014). "Optimization of a free water elimination two-compartment model for diffusion tensor imaging", NeuroImage, vol. 103, pp. 323-333.
Jensen, J. H. (2005). "Diffusional kurtosis imaging: The quantification of non-gaussian water diffusion by means of magnetic resonance imaging", Magnetic Resonance in Medicine, vol. 53, no. 6, pp. 1432-1440
Kukull, W. A. (2002). "Dementia and Alzheimer disease incidence: A prospective cohort study", Archives of neurology, vol. 59, no. 11, pp 1737-1746
Özarslan, E. (2013). "Mean apparent propagator (MAP) MRI: A novel diffusion imaging method for mapping tissue microstructure", NeuroImage, vol. 78, pp. 16-32
Preziosa, P. (2023). "NODDI, diffusion tensor microstructural abnormalities and atrophy of brain white matter and gray matter contribute to cognitive impairment in multiple sclerosis", Journal of neurology, vol. 270, no. 2, pp. 810–823
Zhang, H. (2012). "NODDI: Practical in vivo neurite orientation dispersion and density imaging of the human brain", NeuroImage, vol. 61, no. 4, pp. 1000-1016