Generalizable MRI normative modelling to detect age-inappropriate neurodegeneration

Abstract

Background: Determining whether MRI brain scans demonstrate atrophy that is beyond "normal for age" is challenging. Automated measurements of structural metrics in individual brain regions have shown promise as biomarkers of neurodegeneration, yet widely available reference standards that aid interpretation at the individual level are lacking. Normative modelling, enabling standardized "brain charts", represents a significant step in addressing this challenge by generating individualized age- and sex- adjusted centile scores derived from large, aggregated datasets for MRI-derived quantitative metrics.

Methods: Using normative data from 56,173 participants across the life course, we have developed regional cortical thickness and amygdala/hippocampal volume brain charts (adjusted for total intracranial volume) that can be applied at the individual level. At the group level, we investigate whether regional centile scores relate to cognitive performance (mini-mental state examination) and discriminate individuals with neuropathological evidence of Alzheimer's disease (n = 351) from propensity-matched controls from the National Alzheimer's Coordinating Center (NACC) dataset. In addition, we explored the relationships between disease stage, cognition, regional tau deposition and regional centile scores in amyloid-β-PET-positive individuals with Alzheimer's disease dementia (n = 39) and mild cognitive impairment (n = 71) from the Alzheimer's Disease Neuroimaging Initiative-3 (ADNI-3). We then extended this approach to phenotypes of frontotemporal lobar degeneration using the Neuroimaging in Frontotemporal Dementia dataset (n = 113).

Results: We demonstrate BrainChart's application to illustrative individual cases. At the group level, we show that in Alzheimer's disease, regional centile scores from brain charting predicted cognitive performance, temporal lobe tau PET tracer uptake and discriminated disease groups from propensity matched cognitively normal controls in independent cohorts. Distinct patterns of age-inappropriate cortical atrophy were also evident in different clinical phenotypes of frontotemporal lobar degeneration from the Neuroimaging in Frontotemporal Dementia dataset.

Conclusions: Regional centile scores derived from an extensive normative dataset represent a generalizable method for objectively identifying atrophy in neurodegenerative diseases and can be applied to determine neurodegenerative atrophy at the individual level.

Keywords: Alzheimer’s disease; BrainChart; Dementia; Frontotemporal lobar degeneration; MRI; Neurodegeneration; Normative modelling.

Fig. 1

Brain phenotypes were measured in a reference dataset of MRI scans. Generalized Additive Models for Location, Scale, and Shape (GAMLSS) were used to estimate the relationship between MRI structural metrics and age, stratified by sex, and adjusted for technical and other sources of variation across scanning sites and studies. The normative trajectory, showing the median and confidence intervals for each phenotype is plotted as a population reference curve. Out-of-sample data from a new MRI study are aligned to the corresponding age range of this trajectory using maximum likelihood estimation to estimate study-specific offsets (random effects) for the mean (μ), variance (σ), and skewness (ν) of the statistical distribution, considering age and sex. This alignment allows centile scores to be calculated for each scan on the same scale as the reference population, accounting for study-specific batch effects. Reproduced from Bethlehem, Seidlitz, White et al., Brain charts for the human lifespan [30]. Reproduced under a Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/)

Fig. 2

Ilustrative examples of regional BrainChart centile scores applied to the individual level. Regional cortical thickness centile scores are displayed on an inflated brain and in tabular format (cortical regions arranged by cortical lobe - https://surfer.nmr.mgh.harvard.edu/fswiki/CorticalParcellation). A A 62-year-old male with young onset amnestic Alzheimer’s disease. A coronal slice from a T1 weighted MRI scan and hippocampal volume centile scores are shown. B A 59-year-old female with posterior cortical atrophy. Top row shows sagittal slice of T1 weighted MRI scan and visuospatial tasks from the Addenbrooke's Cognitive Examination-III. C 74-year-old female diagnosed with logopenic variant of primary progressive aphasia. D A 60-year-old male with subjective cognitive complaints, normal cognitive testing performance and no evidence of Alzheimer’s disease on blood and CSF biomarker testing. Key: AD = Alzheimer’s disease

Fig. 3

BrainChart median cortical thickness, hippocampal and amygdala volume centile scores derived from baseline MRI brain scans in pathologically confirmed Alzheimer’s disease (A) (moderate or frequent CERAD neuritic plaque density + Braak stage III-VI); n = 351; Median age = 78.3 years, Median MMSE score = 24/30; Female = 43.3%; Median time to death = 5.9 years. B Spearman’s correlation between BrainChart cortical thickness, hippocampal and amygdala volume centile scores and MMSE score. Only p-values that were significant following FDR correction are shown. Key: AD = Alzheimer’s disease, FDR = False Discovery Rate, MMSE = Mini-mental state examination, NACC = National Alzheimer's Coordinating Center

Fig. 4

Scatter plots highlighting relationship between BrainChart centile scores and MMSE score in Alzheimer’s disease signature regions (a priori regions of interest—red) and primary visual cortex (control region – blue) in pathologically confirmed Alzheimer’s disease (moderate or frequent CERAD neuritic plaque density + Braak stage III-VI; n = 351; Median age = 78.3 years, Median MMSE score = 24/30; Female = 43.3%; Median time to death = 5.9 years). Key: MMSE = Mini-mental state examination

Fig. 5

A Relationship between age and raw uncorrected estimates of cortical thickness, amygdala volume and hippocampal volume in 1400 cognitively normal individuals (median age = 70.5 years, Median MMSE score = 29/30; Female = 66.4%). B Ability of BrainChart centile scores for AD signature regions to discriminate between 351 pathologically confirmed AD cases (moderate or frequent CERAD neuritic plaque density + Braak stage III-VI; Median age = 78.3 years; Median MMSE score = 24/30; Female = 43.3%; Median time to death = 5.9 years) and 351 propensity-matched cognitively normal controls with and without conversion to BrainChart regional centile scores (AUC = area under curve). Centile scores for a priori regions of interest making up the combined model include hippocampal volume, as well as thickness entorhinal, inferior temporal, middle temporal, precuneus and inferior parietal cortices. Key: AD = Alzheimer’s disease, AUC = Area under the curve, FDR = False Discovery Rate, MMSE = Mini-mental state examination, NACC = National Alzheimer's Coordinating Center

Fig. 6

A BrainChart median cortical thickness and hippocampal centile scores derived from baseline MRI brain scans in Aβ-positive MCI (n = 71; Median age = 74.7 years; Median MMSE = 27/30; Female = 53.5%; Median Centiloid scale score = 82); B BrainChart median cortical thickness and hippocampal centile scores derived from baseline MRI brain scans Aβ-positive AD dementia (n = 39; Median age = 75.0 years; Median MMSE = 23/30; Female = 35.9%; Median Centiloid scale score = 96). C Spearman’s correlation between BrainChart cortical thickness and hippocampal volume centile scores and MMSE score at baseline in Aβ-positive MCI and AD dementia in ADNI-3 (n = 110; Median age = 74.8 years; Median MMSE = 27/30; Female = 47.3% Median Centiloid scale score = 87). D Ability of BrainChart centile scores for AD signature regions to discriminate between 71 Aβ-positive MCI cases and 71 Aβ-negative propensity-matched cognitively normal participants. E Ability of BrainChart centile scores for AD signature regions to discriminate between 39 Aβ-positive AD dementia cases and 39 Aβ-negative propensity-matched cognitively normal participants. Centile scores for a priori regions of interest making up the combined model include hippocampal volume, as well as thickness entorhinal, inferior temporal, middle temporal, precuneus and inferior parietal cortices. Key: Aβ = beta-amyloid, AD = Alzheimer’s disease, ADNI = Alzheimer's Disease Neuroimaging Initiative, FDR = False Discovery Rate, MMSE = Mini-mental state examination

Fig. 7

A Spearman’s correlation between BrainChart centile scores and regional tau PET (PET ligand = flortaucipir) SUVR (partial-volume-corrected with the inferior cerebellar grey matter as the reference region) at baseline in Aβ-positive MCI and AD dementia cases in ADNI-3 (n = 97). NB Hippocampus not examined due to flortaucipir signal in hippocampus being contaminated by off-target binding in the choroid plexus [38]. B. Scatter plots between highlighting relationship between BrainChart centile scores and Tau PET SUVR in Alzheimer’s disease signature regions (purple) and primary visual cortex (control region – green). Key: Aβ = beta-amyloid, AD = Alzheimer’s disease, ADNI = Alzheimer's Disease Neuroimaging Initiative, FDR = False Discovery Rate, MCI = Mild Cognitive Impairment, SUVR = standard uptake value ratio for individual regions of interest derived following intensity normalisation using inferior cerebellar grey matter as a reference region

Fig. 8

A BrainChart median cortical centile scores derived from baseline MRI brain scans in patients with behavioural variant frontotemporal dementia (n = 49; Median age = 61 years; Female = 36.7%); B BrainChart median cortical centile scores derived from baseline MRI brain scans in patients with primary progressive aphasia semantic variant (n = 32; Median age = 63.5 years; Female = 43.8%); C BrainChart median cortical centile scores derived from baseline MRI brain scans in patients with primary progressive aphasia non-fluent variant (n = 32; Median age = 68 years; Female = 56.3%)