Improved Sensitivity and Precision in Multicentre Diffusion MRI Network Analysis Using Thresholding and Harmonization

Bruno Miguel de Brito Robalo, Alberto De Luca, Christopher Chen, Anna Dewenter, Marco Duering, Saima Hilal, Huiberdina L. Koek, Anna Kopczak, Bonnie Lam, Alexander Leemans, Vincent Mok, Laurien Onkenhout, Hilde van den Brink, Geert Jan BiesselsSSRN 2022 · DOI

TL;DR

This paper tackles the challenge of analyzing brain networks from multi-site diffusion MRI (dMRI) data, which is often plagued by inconsistencies. By combining network thresholding and data harmonization techniques, the researchers demonstrate a significant improvement in the reliability, precision, and sensitivity of detecting disease-related changes in brain connectivity, specifically in the context of cerebral small vessel disease (SVD).

What Problem Does This Paper Address?

Analyzing brain networks using dMRI is a powerful tool for understanding neurological disorders. However, a major hurdle in this field is the variability and inconsistency of network analyses, especially when combining data from different research sites (multicenter studies). This inconsistency arises from several factors:

  • False-positive connections: dMRI tractography can generate spurious connections, leading to noisy network representations.
  • Acquisition variability: Differences in MRI scanners, protocols, and settings across sites introduce systematic biases in the data.
  • Limited sample sizes: Individual studies often have small sample sizes, making it difficult to detect subtle disease effects and leading to issues with statistical power and multiple comparisons.

These issues hinder the generalizability of findings and the ability to leverage large, existing multicenter datasets to gain a deeper understanding of brain disorders. The authors aim to address these limitations by proposing and evaluating methods to improve the consistency and analytical power of dMRI-based brain network analysis in multicenter settings.

Methods

The study employs two primary techniques to enhance dMRI network analysis:

  • Network Thresholding: This involves filtering out connections with low probability (i.e., those reconstructed in only a small proportion of subjects). This process aims to retain only the more reliable connections, creating a more consistent network architecture. The authors applied a fixed network density of 15% to ensure comparability across sites.
  • Data Harmonization: This technique addresses the issue of scanner-related variability. The researchers used a rotation invariant spherical harmonic (RISH) feature-based harmonization method. This involves using a small set of age and sex-matched controls from each site to scale the raw dMRI signal, aiming to minimize acquisition-related differences while preserving biological effects. This harmonization was applied to the raw dMRI signal before network reconstruction.

These methods were applied individually and in combination to dMRI data from multiple cohorts. The analysis focused on both the binary structure (presence/absence of connections) and the weighted structure (fractional anisotropy - FA - of connections).

Datasets

The study utilized dMRI data from five cohorts with cerebral small vessel disease (SVD) and their respective controls:

  • Utrecht: A memory clinic cohort with SVD patients and a community-based control group.
  • Utrecht 2 (ZOOM): A prospective observational cohort study of SVD patients and controls.
  • Hong Kong: A community-based cohort with SVD patients and controls.
  • Singapore: A community-based cohort with SVD patients and controls.
  • Munich: A cohort of patients with monogenic SVD (CADASIL) and controls, used as a reference standard.

In total, the study included 629 patients and 166 controls. The CADASIL cohort served as a "gold standard" for identifying SVD-related network injury patterns due to its well-defined and severe pathology.

Key Insights and Ideas

  • Thresholding is crucial for consistency: Simply averaging connectivity across sites without filtering leads to highly variable network architectures, with many unreliable connections. Thresholding effectively removes these spurious connections, leading to more consistent network structures across different sites.
  • Harmonization is key for pooling: While thresholding improves consistency, harmonization is essential for enabling the pooling of data from different sites. Without harmonization, differences in FA-weighted connections across sites remain substantial, even after thresholding.
  • The synergy of thresholding and harmonization: The combination of both techniques yields the most robust results, creating networks with both consistent architecture and comparable connection weights across sites. This is vital for increasing statistical power.
  • CADASIL as a valuable reference: Using a genetically defined SVD cohort (CADASIL) as a reference standard provides a crucial benchmark for evaluating the precision and sensitivity of network analyses in sporadic SVD cohorts.
  • The importance of large sample sizes: Individual sites often lack the statistical power to detect subtle disease effects. Harmonization and pooling enable the creation of larger datasets, significantly improving the sensitivity to detect these effects.

Key Results

  • Improved Cross-Site Consistency:
    • Before thresholding and harmonization, cross-site differences in connection probability were substantial (34.7%-46.8%).
  • After thresholding, these differences were significantly reduced (4.5%-7.8%).
  • Harmonization alone did not significantly improve consistency of connection probability.
  • For FA-weighted networks, harmonization alone reduced cross-site differences from up to 36.9% to 10.8%-20.5%.
  • The combination of thresholding and harmonization minimized cross-site differences in both connection probability and FA.
  • Enhanced Precision and Sensitivity in SVD Analysis:
    • Precision: Thresholding dramatically improved precision in detecting SVD-affected connections relative to the CADASIL reference, increasing it from 0.09-0.19 to 0.38-0.70 in individual sporadic SVD sites.
  • Sensitivity: Individual sites showed low sensitivity, with few connections surviving multiple testing correction.
  • Pooled Data: Harmonization and pooling of data significantly improved sensitivity, with 38 connections surviving correction in the harmonized pooled data compared to very few in individual sites.
  • Combined Approach: The combination of thresholding and harmonization in pooled data achieved high precision (0.97) and improved sensitivity.

Limitations and Future Work

  • Lack of a Ground-Truth: The study acknowledges the absence of a definitive "ground-truth" for SVD-related network injury. The CADASIL cohort was used as a model, but it represents a typical example of severe SVD rather than an absolute truth.
  • Patient Selection: The study focused on white matter lesion burden for patient selection and did not harmonize for specific disease stages, risk profiles, or outcomes. Future work could explore these aspects.
  • Technical Choices: The study acknowledges that network reconstruction involves many methodological choices (e.g., registration, tractography algorithms, parcellation schemes) that were not exhaustively tested.
  • Future Directions:
    • Further characterize actual patterns of disrupted connections in SVD and their functional impact using the demonstrated techniques.
  • Utilize machine learning and classification algorithms with large, harmonized multicenter datasets to predict cognitive decline or identify tracts that predict cognitive changes.
  • Investigate topological metrics derived from graph theory in addition to connection probability and weights.
  • Explore the generalizability of these techniques to other neurological diseases.

Summary updated

BibTeX
@misc{robalo2022improved,
  author = {Bruno Miguel de Brito Robalo and Alberto De Luca and Christopher Chen and Anna Dewenter and Marco Duering and Saima Hilal and Huiberdina L. Koek and Anna Kopczak and Bonnie Lam and Alexander Leemans and Vincent Mok and Laurien Onkenhout and Hilde van den Brink and Geert Jan Biessels},
  title = {{Improved Sensitivity and Precision in Multicentre Diffusion MRI Network Analysis Using Thresholding and Harmonization}},
  year = {2022},
  doi = {10.2139/ssrn.4077955},
  howpublished = {SSRN},
}
Figure