Multimodal Brain Network Jointly Construction and Fusion for Diagnosis of Epilepsy

Qi Zhu, Jing Yang, Bingliang Xu, Zhenghua Hou, Liang Sun, Daoqiang ZhangFrontiers in Neuroscience 2021 · DOI

Abstract

Brain network analysis is a cornerstone of modern neurological disease diagnosis. While combining functional and structural imaging data has shown promise, most existing methods construct these networks independently, missing the opportunity to leverage their inherent biological relationships. This paper introduces a unified brain network construction algorithm that jointly learns from functional (fMRI) and structural (DTI) data. By integrating space alignment, low-rank representation for global connectivity, and structural manifold constraints, the proposed framework captures more discriminative features. Furthermore, the PageRank algorithm is employed to identify critical brain regions, and a multi-kernel strategy fuses connectivity and node information for classification. Applied to epilepsy diagnosis, this method demonstrates superior performance over traditional approaches.

Introduction: The Challenge of Multi-modal Fusion

In neuroimaging, researchers typically analyze the brain through two distinct lenses: functional connectivity (FC), derived from synchronized activity in fMRI, and structural connectivity (SC), which maps the physical white matter pathways via DTI. Traditionally, these modalities are processed separately and combined only at the final classification stage—a strategy known as "late fusion."

Conceptual frameworks for multi-modal brain network analysis. Panel C represents the proposed unified approach.

As shown in Figure 1, standard approaches (A) or simple assistive approaches (B) fail to fully exploit the complementary nature of brain structure and function. The physical "wiring" of the brain constrains its functional activity; ignoring this relationship during network construction leads to less biologically accurate models.

Methods: The UM2BN Framework

The researchers propose the Unified Multi-modal Brain Network (UM2BN) framework, which moves beyond simple fusion to joint construction. The workflow is divided into three primary stages:

1. Unified Network Construction

The core innovation lies in a single optimization framework that processes fMRI and DTI data simultaneously.

  • Space Alignment: High-dimensional fMRI data is projected into a lower-dimensional space to reduce noise and computational complexity.
  • Low-Rank Representation: Instead of simple pairwise correlations, the model uses low-rank representation to capture global relationships across all brain regions.
  • Structural Manifold Constraint: DTI data is embedded as a local manifold constraint. This ensures that the functional network construction is guided by the underlying physical anatomy, assuming that regions with strong structural ties should exhibit related functional patterns.
Detailed workflow of the proposed UM2BN method, from data input to classification.

2. Node Importance via PageRank

Beyond just looking at the "edges" (connections) between regions, the framework evaluates the "nodes" (brain regions). The authors adapt the PageRank algorithm—originally designed to rank web pages—to identify the most influential hubs in the brain network. This captures high-order topological information that standard metrics often overlook.

3. Multi-Kernel Fusion

To resolve data heterogeneity, a multi-kernel SVM strategy is employed. This allows the model to optimally weight different feature types—connectivity matrices and node importance scores—to maximize diagnostic accuracy.

Experimental Results

The study evaluated the UM2BN method on a dataset of 306 subjects, including Normal Controls (NC), patients with Frontal Lobe Epilepsy (FLE), and patients with Temporal Lobe Epilepsy (TLE). The proposed method consistently outperformed state-of-the-art techniques such as standard Pearson Correlation (PC) and Graph Convolutional Neural Networks (GCNN).

Classification Accuracy Highlights: * NC vs. FLE: 73.3% * NC vs. TLE: 75.5% * NC vs. (FLE & TLE): 75.0% * FLE vs. TLE: 67.9%

Ablation studies highlighted the necessity of the unified approach: removing the DTI structural constraint or the space alignment step resulted in significant performance drops (between 1.4% and 4.4%).

Discussion and Key Insights

The success of UM2BN underscores several critical points in neuroimaging research:

  • Structure-Function Synergy: Treating DTI data as a mathematical constraint rather than just another feature set produces more discriminative biomarkers.
  • Global Context: Low-rank representation allows the model to consider a region's relationship to the entire brain, which is vital for systemic disorders like epilepsy.
  • Topological Biomarkers: The use of PageRank identified specific regions, such as the Parahippocampal gyrus and Precuneus, as key areas of alteration in epilepsy patients, providing potential targets for clinical focus.

Conclusion

The UM2BN framework represents a significant step forward in multi-modal brain analysis. By jointly constructing networks that respect both functional dynamics and structural constraints, the method provides a more robust and accurate tool for diagnosing epilepsy. Future work will explore more granular brain templates and the application of this framework to other neurological conditions.

Summary updated

Abstract

Brain network analysis has been proved to be one of the most effective methods in brain disease diagnosis. In order to construct discriminative brain networks and improve the performance of disease diagnosis, many machine learning-based methods have been proposed. Recent studies show that combining functional and structural brain networks is more effective than using only single modality data. However, in the most of existing multi-modal brain network analysis methods, it is a common strategy that constructs functional and structural network separately, which is difficult to embed complementary information of different modalities of brain network. To address this issue, we propose a unified brain network construction algorithm, which jointly learns both functional and structural data and effectively face the connectivity and node features for improving classification. First, we conduct space alignment and brain network construction under a unified framework, and then build the correlation model among all brain regions with functional data by low-rank representation so that the global brain region correlation can be captured. Simultaneously, the local manifold with structural data is embedded into this model to preserve the local structural information. Second, the PageRank algorithm is adaptively used to evaluate the significance of different brain regions, in which the interaction of multiple brain regions is considered. Finally, a multi-kernel strategy is utilized to solve the data heterogeneity problem and merge the connectivity as well as node information for classification. We apply the proposed method to the diagnosis of epilepsy, and the experimental results show that our method can achieve a promising performance.
BibTeX
@inproceedings{zhu2021multimodal,
  author = {Qi Zhu and Jing Yang and Bingliang Xu and Zhenghua Hou and Liang Sun and Daoqiang Zhang},
  title = {{Multimodal Brain Network Jointly Construction and Fusion for Diagnosis of Epilepsy}},
  booktitle = {Frontiers in Neuroscience},
  year = {2021},
  doi = {10.3389/fnins.2021.734711},
}
Figure