Synthesizing longitudinal cortical thickness estimates with a flexible and hierarchical multivariate measurement-error model

Jesse W. Birchfield, Nicholas J. Tustison, Andrew J. HolbrookarXiv 2022 · DOI · arXiv

TL;DR

Researchers have developed a Bayesian hierarchical model to solve a major headache in neuroimaging: different software pipelines produce conflicting measurements of brain structures. By treating the "true" cortical thickness as a hidden variable and synthesizing data from seven different pipelines, the model produces an "idealized" estimate that is a significantly more powerful predictor of Alzheimer’s-related cognitive decline than any single software tool used alone.

What Problem Does This Paper Address?

The entorhinal cortex (EC) is the "ground zero" for Alzheimer’s Disease (AD), showing atrophy long before clinical symptoms appear. However, measuring entorhinal cortical thickness (eCT) via MRI is notoriously difficult. The difference between a healthy brain and one with early-stage AD is often a matter of just 1 to 2 millimeters.

Currently, researchers use various software "pipelines" (like FreeSurfer or ANTs) to extract these measurements. The problem? These pipelines often disagree—sometimes by more than the actual physiological difference they are trying to measure. Because we cannot physically measure the brain of a living patient to verify the numbers, there is no "gold standard" or ground truth. Researchers are left guessing which pipeline is most accurate, or simply picking one and ignoring the others.

Methods

The authors propose a Bayesian hierarchical multivariate measurement-error model. Instead of choosing one pipeline, they treat the "true" eCT as a latent (unobserved) variable that informs all seven pipeline outputs simultaneously.

The core technical innovation is the use of a Non-Elliptically Contoured t-distribution (NECT) to model measurement errors. Standard statistical models often assume errors follow a "normal" bell curve. However, MRI data is messy and prone to outliers. The NECT distribution allows the model to account for:

  • Constant Offsets: Some pipelines systematically over- or under-estimate thickness.
  • Differential Tailedness: Some pipelines are generally precise but occasionally produce wild "spikes" (heavy tails), while others are consistently "noisier" but have fewer extreme outliers.
  • Correlation Structures: Pipelines within the same software family (e.g., two versions of FreeSurfer) are likely to make similar mistakes; the model accounts for this so it doesn't "double-count" redundant information.

The model was implemented using Stan, a probabilistic programming language that uses Hamiltonian Monte Carlo to estimate thousands of parameters at once.

Datasets

The study utilized data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI-1):

  • Subjects: 663 individuals (197 cognitively normal, 324 mildly cognitively impaired, and 142 with AD).
  • Visits: A total of 2,449 longitudinal visits.
  • Pipelines: Seven distinct eCT extraction methods (2 from FreeSurfer, 5 from ANTs) were applied to every MRI scan.
  • Clinical Outcome: Mini-Mental State Examination (MMSE) scores were used to measure cognitive capacity.

Key Insights and Ideas

  • The Precision-Outlier Tradeoff: The model revealed a fascinating split between software ecosystems. FreeSurfer pipelines generally have a smaller spread of error (higher precision) but are prone to more extreme outliers (heavier tails). ANTs pipelines have a larger overall error spread but are more robust against extreme "spikes."
  • Strength in Unity: Surprisingly, the "combined" model showed a stronger association between cortical thickness and cognitive capacity than any single pipeline. By propagating the uncertainty of the measurements into the clinical model, the researchers actually recovered a "clearer" signal of disease progression.
  • Longitudinal Smoothing: The idealized eCT estimates are much smoother over time than individual pipeline outputs. The model correctly identifies that a sudden 2mm "gain" in brain thickness in an elderly patient is likely measurement error, not biological growth.

Key Results

  • Predictive Power: Each millimeter of "idealized" eCT was associated with a gain of 0.75 points on the MMSE (95% CI: 0.57, 0.94). This association was clearer and more statistically significant in the combined model than in models using single-pipeline data.
  • Error Characterization: The model quantified systematic offsets. For instance, FreeSurfer pipelines (FSCross/FSLong) tended to underestimate thickness by ~1mm compared to the group average, while ANTsCross overestimated it by ~0.92mm.
  • Clinical Significance: The model successfully captured the interaction between diagnosis and time, showing that AD patients decline significantly faster than those with Mild Cognitive Impairment (MCI), with eCT serving as a critical mediating variable.

Limitations and Future Work

  • Pipeline Selection: While seven pipelines were used, they represent only two major "ecosystems." Future work could incorporate even more diverse computational methods.
  • Linear Offsets: The current model assumes that a pipeline's systematic error (offset) is constant. However, the authors note that a pipeline might be more accurate at measuring thick cortex than thin cortex (or vice versa). A future model could allow offsets to vary linearly with the thickness itself.
  • Scaling Up: The authors hope to extend this framework to all 62 brain regions defined in standard anatomical atlases, potentially providing a holistic, high-precision map of how the entire Alzheimer’s brain evolves over time.

Summary updated

Abstract

MRI-based entorhinal cortical thickness (eCT) measurements predict cognitive decline in Alzheimer's disease (AD) with low cost and minimal invasiveness. Two prominent imaging paradigms, FreeSurfer (FS) and Advanced Normalization Tools (ANTs), feature multiple pipelines for extracting region-specific eCT measurements from raw MRI, but the sheer complexity of these pipelines makes it difficult to choose between pipelines, compare results between pipelines, and characterize uncertainty in pipeline estimates. Worse yet, the EC is particularly difficult to image, leading to variations in thickness estimates between pipelines that overwhelm physiologicl variations predictive of AD. We examine the eCT outputs of seven different pipelines on MRIs from the Alzheimer's Disease Neuroimaging Initiative. Because of both theoretical and practical limitations, we have no gold standard by which to evaluate them. Instead, we use a Bayesian hierarchical model to combine the estimates. The resulting posterior distribution yields high-probability idealized eCT values that account for inherent uncertainty through a flexible multivariate error model that supports different constant offsets, standard deviations, tailedness, and correlation structures between pipelines. Our hierarchical model directly relates idealized eCTs to clinical outcomes in a way that propagates eCT estimation uncertainty to clinical estimates while accounting for longitudinal structure in the data. Surprisingly, even though it incorporates greater uncertainty in the predictor and regularization provided by the prior, the combined model reveals a stronger association between eCT and cognitive capacity than do nonhierarchical models based on data from single pipelines alone.
BibTeX
@misc{birchfield2022synthesizing,
  author = {Jesse W. Birchfield and Nicholas J. Tustison and Andrew J. Holbrook},
  title = {{Synthesizing longitudinal cortical thickness estimates with a flexible and hierarchical multivariate measurement-error model}},
  year = {2022},
  eprint = {2204.03576},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2204.03576},
}
Figure