Joint Learning of Multiple Differential Networks with fMRI data for Brain Connectivity Alteration Detection

Hao Chen, Ying Guo, Yong He, Dong Liu, Lei Liu, Xiao‐Hua ZhouarXiv 2021 · DOI · arXiv

Abstract

Identifying brain connectivity alterations is fundamental to understanding neurological disorders. However, multi-center fMRI studies are often hampered by technical heterogeneity across sites, such as differing scanners and imaging protocols. This paper introduces JLMDN (Joint Learning of Multiple Differential Networks), a framework designed to jointly estimate differential networks from multiple datasets. By utilizing a sparse group Minimax Concave Penalty (gMCP), JLMDN captures common pathological structures while accounting for site-specific variations. Experimental results on ADHD-200 datasets demonstrate superior performance in both network identification and disease classification compared to existing methods.

Introduction

In neuroimaging, a "differential network" identifies the specific brain connections that differ between a diseased group and a healthy control group. While multi-center studies provide the large sample sizes necessary for statistical power, they introduce a "heterogeneity problem": different sites use different scanners and imaging protocols, leading to distinct statistical properties. Simply pooling data can bias results, while independent analysis ignores shared disease mechanisms. JLMDN addresses this by jointly learning these networks, allowing researchers to "borrow strength" across sites to uncover shared pathological features.

Methodology

The JLMDN framework follows a three-stage pipeline to transform raw fMRI data into interpretable differential networks.

The JLMDN pipeline: (A) Data preprocessing and individual network estimation, (B) Bootstrap-assisted ensemble learning, and (C) Final differential network identification.

1. Individual Network Estimation

For each subject, a spatial precision matrix—representing direct connectivity between brain regions—is estimated using the CLIME (Constrained $\ell_1$-minimization) method. These partial correlations are then Fisher-transformed to stabilize variance and handle potential outliers.

2. Penalized Logistic Regression

The task is framed as a classification problem where brain connectivity strengths serve as predictors for disease status. This allows the model to identify which edges are most discriminative for the condition being studied.

3. Sparse Group MCP (gMCP)

The core innovation is the application of a Group Minimax Concave Penalty. This penalty induces sparsity at two levels:

  • Group Level: It encourages the selection of the same "differential edges" across all research sites, capturing common disease mechanisms.
  • Individual Level: It allows for site-specific sparsity, accommodating technical variations inherent in multi-center data.

4. Ensemble Learning

To enhance stability and empirical performance, the authors implement a bootstrap-assisted ensemble procedure, aggregating results from multiple data sub-samples to ensure the robustness of the identified networks.

Experimental Validation

The authors validated JLMDN using both synthetic and real-world data. Simulation studies employed "Hub" and "Small-World" network structures to test the model's ability to identify commonalities across varying degrees of overlap. For real-world application, the ADHD-200 Study was utilized, specifically analyzing resting-state fMRI data from Peking University and New York University.

Results and Discussion

JLMDN consistently outperformed state-of-the-art methods in simulations, particularly in terms of True Discovery Rate (TDR) and Precision-Recall curves. In the ADHD-200 analysis, the framework identified key differential regions in the Frontal, Parietal, and Cerebellum areas. These findings align closely with clinical literature regarding executive control and motor learning impairments in ADHD.

Example of a detected differential network, highlighting altered connectivity patterns in the brain.

Furthermore, the JLMDN-based model achieved high classification accuracy in cross-site validation, suggesting that the identified differential edges are highly reliable biomarkers for distinguishing disease states.

Conclusion

JLMDN provides a robust framework for multi-center neuroimaging research. By effectively handling site heterogeneity and small sample sizes, it offers a more accurate way to detect brain connectivity alterations. Future research will focus on identifying "hub nodes"—critical junctions in the brain—and improving computational efficiency for high-resolution brain atlases.

This summary is based on the research paper "Joint Learning of Multiple Differential Networks with fMRI data for Brain Connectivity Alteration Detection."

Summary updated

Abstract

In this study we focus on the problem of joint learning of multiple differential networks with function Magnetic Resonance Imaging (fMRI) data sets from multiple research centers. As the research centers may use different scanners and imaging parameters, joint learning of differential networks with fMRI data from different centers may reflect the underlying mechanism of neurological diseases from different perspectives while capturing the common structures. We transform the task as a penalized logistic regression problem, and exploit sparse group Minimax Concave Penalty (gMCP) to induce common structures among multiple differential networks and the sparse structures of each differential network. To further enhance the empirical performance, we develop an ensemble-learning procedure. We conduct thorough simulation study to assess the finite-sample performance of the proposed method and compare with state-of-the-art alternatives. We apply the proposed method to analyze fMRI datasets related with Attention Deficit Hyperactivity Disorder from various research centers. The identified common hub nodes and differential interaction patterns coincides with the existing experimental studies.
BibTeX
@misc{chen2021joint,
  author = {Hao Chen and Ying Guo and Yong He and Dong Liu and Lei Liu and Xiao‐Hua Zhou},
  title = {{Joint Learning of Multiple Differential Networks with fMRI data for Brain Connectivity Alteration Detection}},
  year = {2021},
  eprint = {2106.03334},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2106.03334},
}
Figure