Altered Brain Morphometry in Cerebral Small Vessel Disease With Cerebral Microbleeds: An Investigation Combining Univariate and Multivariate Pattern Analyses

Jing Li, Hongwei Wen, Shengpei Wang, Yena Che, Nan Zhang, Lingfei GuoFrontiers in Neurology 2022 · DOI

This study investigates how Cerebral Small Vessel Disease (CSVD)—a leading cause of stroke and dementia—alters the physical structure of the brain, specifically when accompanied by Cerebral Microbleeds (CMBs). By combining traditional group-level statistics with machine learning, the researchers achieved over 80% accuracy in identifying patients based on brain scans, revealing that microbleeds are associated with distinct, distributed patterns of white matter loss that traditional methods often miss.

TL;DR

Researchers used a combination of voxel-based morphometry and multivariate pattern analysis (MVPA) to map brain volume changes in CSVD patients with and without microbleeds. Their machine-learning model achieved high accuracy (>80%) in classifying patients and discovered that while group-level differences are subtle, the presence of microbleeds creates a unique "fingerprint" of structural damage in the brain's white matter.

What Problem Does This Paper Address?

Cerebral Small Vessel Disease (CSVD) is a "silent" killer of brain tissue, often manifesting as small infarctions, lesions in white matter, or tiny hemorrhages known as microbleeds (CMBs). While we know CSVD leads to cognitive decline, it is difficult to diagnose at the individual level in its early stages.

Traditional "univariate" analysis looks at one tiny 3D pixel (voxel) at a time across a whole group to find average differences. However, this often fails to capture the complex, interconnected ways the brain changes. The authors sought to bridge this gap by using "multivariate" analysis, which looks at the whole brain as a single pattern, to see if they could distinguish between healthy individuals, CSVD patients without microbleeds, and those with them.

Methods

The study employed two primary analytical pipelines:

  1. Univariate Analysis (VBM-DARTEL): The researchers used Voxel-Based Morphometry (VBM) with the DARTEL algorithm to compare Gray Matter Volume (GMV) and White Matter Volume (WMV) across groups. This is the standard "gold-standard" for finding localized tissue loss.
  2. Multivariate Pattern Analysis (MVPA) with Multiple Kernel Learning (MKL): This is the paper's core innovation. Instead of looking at voxels in isolation, MKL treats different brain regions (defined by the AAL atlas) as "kernels." The model learns the contribution of each region to a classification decision. By combining GMV and WMV data into a single model, the researchers could perform binary classification (e.g., Patient vs. Control).
  3. Cross-Validation: They used a nested "leave-one-subject-out" cross-validation approach to ensure the machine learning results were generalizable and not just "overfitting" to their specific participants.

Datasets

The study involved a total of 108 participants recruited between 2018 and 2019:

  • 26 CSVD-c: Patients with CSVD and at least one cerebral microbleed.
  • 43 CSVD-n: Patients with CSVD but no microbleeds.
  • 39 Healthy Controls (HC): Age- and sex-matched individuals with no history of neurological disease.

All participants underwent high-resolution 3.0T MRI scans and a battery of cognitive tests (MoCA, AVLT, SDMT, etc.).

Key Insights and Ideas

  • The Default Mode Network (DMN) is the Primary Target: Both univariate and multivariate analyses pointed to damage in the frontal gyrus and anterior cingulate. These areas are hubs of the DMN, which is critical for internal thought and memory. This explains why CSVD patients often struggle with executive function and memory long before major strokes occur.
  • Microbleeds Create a Distributed "Fingerprint": Interestingly, the univariate analysis found no significant difference between CSVD patients with microbleeds and those without. However, the multivariate MKL model could distinguish them with 81.16% accuracy. This suggests that microbleeds don't just cause one "hole" in the brain; they lead to subtle, widespread alterations in white matter tracts (like the olfactory cortex and middle frontal gyrus) that only a pattern-recognition algorithm can detect.
  • Multimodal Synergy: The study found that combining gray matter and white matter data (the "GMV+WMV" approach) consistently outperformed using either modality alone. This highlights that CSVD is a "whole-brain" disease affecting both the processing centers (gray matter) and the wiring (white matter).

Key Results

  • Classification Accuracy:
    • CSVD-c vs. Healthy Controls: 86.15% accuracy.
  • CSVD-n vs. Healthy Controls: 81.71% accuracy.
  • CSVD-c vs. CSVD-n: 81.16% accuracy.
  • Cognitive Impact: Patients with microbleeds (CSVD-c) performed significantly worse on almost all cognitive tests compared to both the CSVD-n group and healthy controls, particularly in processing speed (SDMT) and executive function (TMT).
  • Structural Loss: Significant volume reduction was found in the right superior frontal gyrus and left anterior cingulate gyrus in all CSVD patients compared to controls.

Limitations and Future Work

  • Sample Size: While 108 participants is respectable for neuroimaging, it is relatively small for complex machine learning. Larger cohorts are needed to validate these "biomarkers."
  • Cross-Sectional Design: The study captures a snapshot in time. It cannot definitively say if the structural changes cause the microbleeds or if the microbleeds cause the structural changes.
  • Feature Engineering: The authors suggest that future models should incorporate more "non-imaging" features, such as age, vascular risk factors (hypertension, diabetes), and genetic markers (like APOE ε4), to further improve diagnostic precision.
  • Toolbox Evolution: The authors noted that newer toolboxes (like CAT12) might be more sensitive than the VBM8 used here, suggesting a technical path for even more precise future studies.

Summary updated

Abstract

Our findings not only suggested that the localized alterations in GMV and WMV appeared to be associated with the pathophysiology of CSVD but also indicated that altered brain morphometry could be a potential discriminative pattern to detect CSVD at the individual level.
BibTeX
@inproceedings{li2022altered,
  author = {Jing Li and Hongwei Wen and Shengpei Wang and Yena Che and Nan Zhang and Lingfei Guo},
  title = {{Altered Brain Morphometry in Cerebral Small Vessel Disease With Cerebral Microbleeds: An Investigation Combining Univariate and Multivariate Pattern Analyses}},
  booktitle = {Frontiers in Neurology},
  year = {2022},
  doi = {10.3389/fneur.2022.819055},
}
Figure