Voxel-wise normative models for abnormality-driven contrast of microstructural MRI

2406 

Abstract Submission 

Olivier Parent¹, Yohan Yee², Gabriel Devenyi², Aurélie Bussy³, Grace Pigeau³, Manuela Costantino³, Jérémie Fouquet¹, Daniela Quesada-Rodriguez³, Mahsa Dadar³, Mallar Chakravarty⁴

¹Douglas Mental Health University Institute, Montreal, QC, ²McGill University, Montreal, Quebec, ³McGill University, Montreal, QC, ⁴Brain Imaging Centre, Douglas Research Centre, Montreal, Quebec

Olivier Parent  
Douglas Mental Health University Institute
Montreal, QC

Yohan Yee  
McGill University
Montreal, Quebec

Gabriel Devenyi  
McGill University
Montreal, Quebec

Aurélie Bussy  
McGill University
Montreal, QC

Grace Pigeau  
McGill University
Montreal, QC

Manuela Costantino  
McGill University
Montreal, QC

Jérémie Fouquet  
Douglas Mental Health University Institute
Montreal, QC

Daniela Quesada-Rodriguez  
McGill University
Montreal, QC

Mahsa Dadar  
McGill University
Montreal, QC

Mallar Chakravarty, PhD  
Brain Imaging Centre, Douglas Research Centre
Montreal, Quebec

Microstructural magnetic resonance imaging (MRI) can be used to measure brain tissue properties in vivo and has been used extensively in scientific research to derive biomarkers sensitive to disease pathology and progression, like normal-appearing white matter alterations in multiple sclerosis (MS) [1] and small vessel disease (SVD) [2] and iron accumulation in the subcortical grey matter in Parkinson's disease (PD) [3]. However, these biomarkers have been seldom implemented in clinical practice. We posit that this limited clinical translation is due, in part, to the difficult interpretability of raw microstructural maps to detect subtle pathological tissue alterations distinct from normal variations. For example, white matter deterioration and subcortical grey matter iron accumulation are both normal age-related phenomena [4,5]. Here, we constructed voxel-wise normative models of microstructural MRI and produced subject-wise z-scored maps of microstructural abnormality relative to these age- and sex-specific averages. These maps bring important context regarding normal brain tissue integrity and permit intuitive visual assessment of brain tissue abnormalities at first glance.

We used data from 32,935 UK Biobank (UKB) subjects (32,014 healthy subjects and 921 subjects with a neurological diagnosis). T1-weighted (T1w) and Fluid-Attenuated Inversion Recovery (FLAIR) were used to segment the brain into 9 broad regions (labels) using the Brain tISsue segmentatiON (BISON) pipeline [6] (Fig. 1A). Diffusion-weighted images were used to derive diffusion tensor imaging (DTI) and neurite orientation dispersion and density imaging (NODDI) microstructural metrics, while susceptibility-weighted images were used to derive T2* and quantitative susceptibility mapping (QSM) metrics (resampled at 2 mm isotropic to match the diffusion resolution) [7]. Multispectral registration to a custom unbiased UKB template [8] was performed using T1w and fractional anisotropy (FA) maps as inputs in order to align tissue boundaries and white matter fibers, respectively. In this common space, we calculated voxel-wise and region-specific normative models using Bayesian linear regression with sex and age modelled with 4th-order B-splines with the PCN toolkit [9], using only healthy subject data. Importantly, in voxels where different regions are present across subjects due to imperfect registration, different normative models were fitted (Fig. 1B). Normative models were only calculated when the label prevalence was higher than 100 subjects (Fig. 1C). This process resulted in region-, age-, and sex-specific atlases of microstructure, which are then used to z-score subject microstructural maps (Fig. 1D). Microstructural maps were denoised prior to the z-scoring process [10].

We present five case studies comparing raw and z-scored microstructural maps in common space (Fig. 2). Comparing two healthy subjects (Fig. 2A), the raw microstructural maps look highly similar (top rows) while z-scored maps reveal subvisible alterations in the white matter (bottom rows) for the second subject, potentially indicating SVD. Comparing two subjects with MS (Fig. 2B), z-scored maps reveal widespread microstructural alterations in white matter for the first subject, while being contained to lesions for the second subject. Lastly, z-scored maps for a subject with PD (Fig. 2C) reveal alterations in the subcortical grey matter, potentially indicating iron accumulation.

Using normative models, we developed a visualization technique that simplifies tissue abnormality detection at the single-subject level. This technique permits a highly intuitive assessment of brain health and could help translate biomarker findings related to brain microstructure from research to clinical settings. Future work will focus on harmonization across sites and extending this principle to voxel-wise volumetric measurements.

Aging

Bayesian Modeling

Methods Development ²

Multi-Modal Imaging ¹

Aging

Data analysis

DISORDERS

Modeling

Statistical Methods