Microstructural mapping of neural pathways in Alzheimer's disease using macrostructure-informed normative tractometry

Abstract

Introduction: Diffusion-weighted magnetic resonance imaging (dMRI) is sensitive to the microstructural properties of brain tissues and shows great promise in detecting the effects of degenerative diseases. However, many approaches analyze single measures averaged over regions of interest without considering the underlying fiber geometry.

Methods: We propose a novel macrostructure-informed normative tractometry (MINT) framework to investigate how white matter (WM) microstructure and macrostructure are jointly altered in mild cognitive impairment (MCI) and dementia. We compared MINT-derived metrics with univariate diffusion tensor imaging (DTI) metrics to examine how fiber geometry may impact the interpretation of microstructure.

Results: In two multisite cohorts from North America and India, we find consistent patterns of microstructural and macrostructural anomalies implicated in MCI and dementia; we also rank diffusion metrics' sensitivity to dementia.

Discussion: We show that MINT, by jointly modeling tract shape and microstructure, has the potential to disentangle and better interpret the effects of degenerative disease on the brain's neural pathways.

Highlights: Changes in diffusion tensor imaging metrics may be due to macroscopic changes. Normative models encode normal variability of diffusion metrics in healthy controls. Variational autoencoder applied on tractography can learn patterns of fiber geometry. WM microstructure and macrostructure are modeled with multivariate methods. Transfer learning uses pretraining and fine-tuning for increased efficiency.

Keywords: Alzheimer's disease; anomaly detection; deep generative models; diffusion magnetic resonance imaging; normative modeling; tractometry; transfer learning.

FIGURE 1

WM bundles from HCP‐842 atlas used in RecoBundles in this study. Bundles visualized individually outside of the glass brain are not to scale. While bilateral association and projection bundles were used in this study, only those in the left hemisphere are shown.

FIGURE 2

Modeling whole‐brain macro‐ and microstructure in MINT. (A) MINT uses 1D convolutional layers in a VAE to model the sequential dependency along each streamline. (B) MINT models whole‐brain fiber geometry by training the VAE on streamlines from multiple fiber bundles. Proximity and similarity between streamlines are encoded using global coordinates. (C) Whole‐brain macrostructure is represented using tractography data, and whole‐brain microstructure is represented by projecting scalar maps of DTI metrics onto every point on each streamline. Both types of data are combined to train the VAE model. (D) The VAE is a generative model, trained via backpropagation, to reconstruct its input data as accurately as possible while passing all the intermediate data through a narrow bottleneck layer (z in middle). VAE enforces the data to be optimally compressed into a latent set of parameters, and the means μ and covariance σ of these parameters follow a multivariate Gaussian distribution. This enables the rich complexity of variations in fiber microstructure and 3D geometry to be modeled using a compact representation. (E) MINT framework. A VAE model is first trained on healthy controls from a pretraining dataset (TractoInferno). For each target dataset (ADNI, NIMHANS), the controls are split into two sets (Split 1 and Split 2), and Split 1 is used to fine‐tune the VAE model. We then perform model inference on all data from the target dataset to obtain their reconstruction for subsequent bundle profiling. For each bundle, the point‐wise deviation score (MAE) is indexed using 100 along‐tract segments created to align data across subjects. MAE for all points within each segment is averaged to create a bundle profile of deviation scores. The BUAN statistical framework is used for harmonization and group difference testing. In the harmonization step, bundle profiles created from data in Split 1 are used to train ComBat,, whose parameters are applied to bundle profiles from healthy controls in Split 2 and the cases (participants with MCI and AD). Harmonized bundle profiles are then used for group difference testing using linear regression. AD, Alzheimer's disease; ADNI, Alzheimer's Disease Neuroimaging Initiative; MCI, mild cognitive impairment; NIMHANS, National Institute of Mental Health and Neurosciences; BUAN, bundle analytics; DTI, diffusion tensor imaging; MAE, mean absolute error; MINT, macrostructure‐informed normative tractometry; VAE, variational autoencoder.

FIGURE 3

Effects of dementia and MCI on MD. Here we show along‐tract effects of diagnostic group by color‐coding the regression coefficients associated with diagnostic group, β(DX), for DTI‐MD and MAE‐MD, for dementia versus CN and MCI versus CN in the ADNI and NIMHANS cohorts, categorized by WM pathways. The colors are mapped using different scales for MAE and DTI values, as MAE measures – calculated from reconstruction error – generally are smaller in magnitude than the DTI measures. The plots show patterns of effects but are not intended for direct comparison of effect sizes across measures, except across three corresponding diffusivity measures where the color map range is the same. ADNI, Alzheimer's Disease Neuroimaging Initiative; CN, cognitively normal; DTI‐MD, diffusion tensor imaging‐mean diffusivity; MAE, mean absolute error; MCI, mild cognitive impairment; MD, mean diffusivity; NIMHANS, National Institute of Mental Health and Neuro Sciences; WM, white matter.

FIGURE 4

Effects of dementia and MCI on FA. Along‐tract β(DX) for DTI‐FA and MAE‐FA, for dementia versus CN and MCI versus CN in the ADNI and NIMHANS cohorts, categorized by WM pathways. The color map is flipped for DTI‐FA compared to MAE‐FA and MAE‐Shape in Figure 5, to better illustrate disease effects on these measures (red). ADNI, Alzheimer's Disease Neuroimaging Initiative; CN, cognitively normal; DTI‐FA, diffusion tensor imaging‐fractional anisotropy; FA, fractional anisotropy; MAE‐FA, mean absolute error‐fractional anisotropy; MCI, mild cognitive impairment; NIMHANS, National Institute of Mental Health and Neurosciences; WM, white matter.

FIGURE 5

Effects of dementia and MCI on MAE‐Shape. Along‐tract β(DX) for MAE‐Shape, for dementia versus CN and MCI versus CN in the ADNI and NIMHANS cohorts, categorized by WM pathways. ADNI, Alzheimer's Disease Neuroimaging Initiative; CN, cognitively normal; MAE, mean absolute error; MCI, mild cognitive impairment; NIMHANS, National Institute of Mental Health and Neurosciences; WM, white matter.

FIGURE 6

Partial Spearman rank‐order correlation coefficient (ρ) and its 95% confidence interval for all 30 bundles between nine diffusion measures and CDRsb (A) and WLDR (B), calculated in Section 2.3.3. The corresponding DTI and MAE measures for each anisotropy and diffusivity metric are shown in the same panel, and MAE‐Shape is shown separately. Correlation coefficients that are statistically significant after FDR correction are marked with a star to the right of each plot with the corresponding color. CDRsb, Clinical Dementia Rating Sum of Boxes; DTI, diffusion tensor imaging; FDR, false discovery rate; MAE, mean absolute error; WLDR, word list delayed recall.

FIGURE 7

Ranking diffusion metrics for their sensitivity to dementia. Here we show a Q‐Q plot of −log₁₀(P) after FDR correction for dementia versus CN group comparison in ADNI cohort, for four DTI‐ and five MINT‐derived diffusion measures in ILF_L, OPT_L, and CC_ForcepsMajor. Microstructural metrics that are associated most strongly with dementia are typically MD and RD, shown in orange and green colors. The dashed line in these plots shows the pattern of p values that would be expected for null measures, that is, brain metric that shows no association with dementia. For some tracts, the MAE‐FA is not much better than a null measure (ie, shows no detectable association with dementia). ADNI = Alzheimer's Disease Neuroimaging Initiative. CN, cognitively normal; DTI, diffusion tensor imaging; FDR, false discovery rate; MAE‐FA, mean absolute error‐fractional anisotropy; MINT, macrostructure‐informed normative tractometry; QQ, quantile‐quantile; RD, radial diffusivity.