Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MvHo-IB: Multi-View Higher-Order Information Bottleneck for Brain Disorder Diagnosis

Kunyu Zhang, Qiang Li, Shujian Yu

TL;DR

The paper addresses brain-disorder diagnosis from rs-fMRI by capturing higher-order interactions among brain regions, going beyond traditional pairwise connectivity. It introduces MvHo-IB, a multi-view information bottleneck framework that combines pairwise functional connectivity and HOIs via $\mathcal{O}$-information, with a matrix-based Rényi entropy estimator and a Brain3DCNN encoder. The learning objective maximizes $I(Y; Z) - (\beta_1 I(X_1; Z_1) + \beta_2 I(X_2; Z_2))$, which for deterministic encoders reduces to ${\rm CE}(Y, \hat{Y}) + \beta_1 H(Z_1) + \beta_2 H(Z_2)$, and a fusion module $Z=f_\theta(Z_1, Z_2)$. Across three benchmark fMRI datasets, MvHo-IB achieves state-of-the-art accuracy, provides neurobiologically plausible three-way HOI biomarkers via Grad-CAM, and offers a scalable route toward higher-order HOIs with potential clinical impact.

Abstract

Recent evidence suggests that modeling higher-order interactions (HOIs) in functional magnetic resonance imaging (fMRI) data can enhance the diagnostic accuracy of machine learning systems. However, effectively extracting and utilizing HOIs remains a significant challenge. In this work, we propose MvHo-IB, a novel multi-view learning framework that integrates both pairwise interactions and HOIs for diagnostic decision-making, while automatically compressing task-irrelevant redundant information. MvHo-IB introduces several key innovations: (1) a principled method that combines O-information from information theory with a matrix-based Renyi alpha-order entropy estimator to quantify and extract HOIs, (2) a purpose-built Brain3DCNN encoder to effectively utilize these interactions, and (3) a new multi-view learning information bottleneck objective to enhance representation learning. Experiments on three benchmark fMRI datasets demonstrate that MvHo-IB achieves state-of-the-art performance, significantly outperforming previous methods, including recent hypergraph-based techniques. The implementation of MvHo-IB is available at https://github.com/zky04/MvHo-IB.

MvHo-IB: Multi-View Higher-Order Information Bottleneck for Brain Disorder Diagnosis

TL;DR

The paper addresses brain-disorder diagnosis from rs-fMRI by capturing higher-order interactions among brain regions, going beyond traditional pairwise connectivity. It introduces MvHo-IB, a multi-view information bottleneck framework that combines pairwise functional connectivity and HOIs via -information, with a matrix-based Rényi entropy estimator and a Brain3DCNN encoder. The learning objective maximizes , which for deterministic encoders reduces to , and a fusion module . Across three benchmark fMRI datasets, MvHo-IB achieves state-of-the-art accuracy, provides neurobiologically plausible three-way HOI biomarkers via Grad-CAM, and offers a scalable route toward higher-order HOIs with potential clinical impact.

Abstract

Recent evidence suggests that modeling higher-order interactions (HOIs) in functional magnetic resonance imaging (fMRI) data can enhance the diagnostic accuracy of machine learning systems. However, effectively extracting and utilizing HOIs remains a significant challenge. In this work, we propose MvHo-IB, a novel multi-view learning framework that integrates both pairwise interactions and HOIs for diagnostic decision-making, while automatically compressing task-irrelevant redundant information. MvHo-IB introduces several key innovations: (1) a principled method that combines O-information from information theory with a matrix-based Renyi alpha-order entropy estimator to quantify and extract HOIs, (2) a purpose-built Brain3DCNN encoder to effectively utilize these interactions, and (3) a new multi-view learning information bottleneck objective to enhance representation learning. Experiments on three benchmark fMRI datasets demonstrate that MvHo-IB achieves state-of-the-art performance, significantly outperforming previous methods, including recent hypergraph-based techniques. The implementation of MvHo-IB is available at https://github.com/zky04/MvHo-IB.

Paper Structure

This paper contains 15 sections, 9 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Information Bottleneck in Brain Disorder Diagnosis
  3. Methodology
  4. Multi-view Information Bottleneck
  5. $\mathcal{O}$-Information and Matrix-based Entropy Functional
  6. Brain3DCNN
  7. 3D Edge-to-Edge (E2E) Layer
  8. 3D Edge-to-Node (E2N) Layer
  9. 3D Node-to-Graph (N2G) Layer
  10. Experiments and Results
  11. Datasets and Experimental Settings
  12. Experimental Results and Ablation Study
  13. Conclusions and Future Work
  14. Acknowledgments.
  15. Disclosure of Interests.

Figures (3)

  • Figure 1: Illustration of MvHo-IB. The time courses are extracted from fMRI using the automated anatomical labeling (AAL) template tzourio2002automated. Then, functional connectivity patterns are estimated before being fed into a multi-view framework. The framework learns a joint representation $Z = f_{\theta}(Z_1, Z_2)$ by balancing the maximization of $I(Y; Z)$ with the minimization of $I(X_1; Z_1) + I(X_2; Z_2)$, where the first view input is the mutual information matrix (pairwise interactions) and the second view input is the $\mathcal{O}$-information 3D tensor (triple interactions).
  • Figure 2: Each block represents the input/output of filter layers. The brain network adjacency matrix (leftmost block) undergoes E2E convolution, followed by E2N filtering to aggregate edge weights per region. The N2G layer integrates node responses, and fully connected layers refine features for the final prediction.
  • Figure 3: The top two interpretable pairwise and three-way interactions used by our model, identified with Grad-CAM. For EOEC, the top two pairwise groups are identical, only one group is presented.