Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

BrainOOD: Out-of-distribution Generalizable Brain Network Analysis

Jiaxing Xu, Yongqiang Chen, Xia Dong, Mengcheng Lan, Tiancheng Huang, Qingtian Bian, James Cheng, Yiping Ke

TL;DR

BrainOOD addresses the dual challenges of out-of-distribution generalization and interpretability in brain-network analysis by coupling a learnable feature selector with a discrete-structure extractor, all guided by an enhanced Graph Information Bottleneck. The framework emphasizes recovering causally informative subgraphs while aligning interpretations across subjects, aided by a high-pass GNN for feature reconstruction and an alignment loss to stabilize structure across a batch. Empirical results on ABIDE and ADNI show BrainOOD outperforming 16 baselines with up to 8.5% improvements in OOD accuracy and through-case analyses that yield neuroscience-consistent biomarkers. Additionally, BrainOOD contributes the first OOD brain-network benchmark, enabling standardized evaluation and reproducibility for future studies in clinical neuroscience and network science.

Abstract

In neuroscience, identifying distinct patterns linked to neurological disorders, such as Alzheimer's and Autism, is critical for early diagnosis and effective intervention. Graph Neural Networks (GNNs) have shown promising in analyzing brain networks, but there are two major challenges in using GNNs: (1) distribution shifts in multi-site brain network data, leading to poor Out-of-Distribution (OOD) generalization, and (2) limited interpretability in identifying key brain regions critical to neurological disorders. Existing graph OOD methods, while effective in other domains, struggle with the unique characteristics of brain networks. To bridge these gaps, we introduce BrainOOD, a novel framework tailored for brain networks that enhances GNNs' OOD generalization and interpretability. BrainOOD framework consists of a feature selector and a structure extractor, which incorporates various auxiliary losses including an improved Graph Information Bottleneck (GIB) objective to recover causal subgraphs. By aligning structure selection across brain networks and filtering noisy features, BrainOOD offers reliable interpretations of critical brain regions. Our approach outperforms 16 existing methods and improves generalization to OOD subjects by up to 8.5%. Case studies highlight the scientific validity of the patterns extracted, which aligns with the findings in known neuroscience literature. We also propose the first OOD brain network benchmark, which provides a foundation for future research in this field. Our code is available at https://github.com/AngusMonroe/BrainOOD.

BrainOOD: Out-of-distribution Generalizable Brain Network Analysis

TL;DR

BrainOOD addresses the dual challenges of out-of-distribution generalization and interpretability in brain-network analysis by coupling a learnable feature selector with a discrete-structure extractor, all guided by an enhanced Graph Information Bottleneck. The framework emphasizes recovering causally informative subgraphs while aligning interpretations across subjects, aided by a high-pass GNN for feature reconstruction and an alignment loss to stabilize structure across a batch. Empirical results on ABIDE and ADNI show BrainOOD outperforming 16 baselines with up to 8.5% improvements in OOD accuracy and through-case analyses that yield neuroscience-consistent biomarkers. Additionally, BrainOOD contributes the first OOD brain-network benchmark, enabling standardized evaluation and reproducibility for future studies in clinical neuroscience and network science.

Abstract

In neuroscience, identifying distinct patterns linked to neurological disorders, such as Alzheimer's and Autism, is critical for early diagnosis and effective intervention. Graph Neural Networks (GNNs) have shown promising in analyzing brain networks, but there are two major challenges in using GNNs: (1) distribution shifts in multi-site brain network data, leading to poor Out-of-Distribution (OOD) generalization, and (2) limited interpretability in identifying key brain regions critical to neurological disorders. Existing graph OOD methods, while effective in other domains, struggle with the unique characteristics of brain networks. To bridge these gaps, we introduce BrainOOD, a novel framework tailored for brain networks that enhances GNNs' OOD generalization and interpretability. BrainOOD framework consists of a feature selector and a structure extractor, which incorporates various auxiliary losses including an improved Graph Information Bottleneck (GIB) objective to recover causal subgraphs. By aligning structure selection across brain networks and filtering noisy features, BrainOOD offers reliable interpretations of critical brain regions. Our approach outperforms 16 existing methods and improves generalization to OOD subjects by up to 8.5%. Case studies highlight the scientific validity of the patterns extracted, which aligns with the findings in known neuroscience literature. We also propose the first OOD brain network benchmark, which provides a foundation for future research in this field. Our code is available at https://github.com/AngusMonroe/BrainOOD.

Paper Structure

This paper contains 34 sections, 2 theorems, 15 equations, 7 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Brain Networks Classification
  4. Graph Neural Networks (GNNs)
  5. Out-of-distribution Benchmark in Brain Network Analysis
  6. Distribution Shifts in Brain Network Analysis
  7. Dataset under OOD Setting
  8. BrainOOD
  9. Interpretable and Generalizable Brain Network Analysis
  10. Feature Selection via Reconstruction
  11. Structure Extraction by Discrete Sampling
  12. Loss Functions
  13. Experimental Results
  14. Baseline Models
  15. Main Results
  16. ...and 19 more sections

Key Result

Theorem 4.1

For a subgraph extractor $g_\phi$ that encodes the input graph $G$ into representation ${\bm{H}}$ to extract the desired subgraph $G_C^*$, if $g_\phi$ is limited in representation power, i.e., $I(G;{\bm{H}})< H(G_C^*)$, where $H(\cdot)$ is the entropy of the underlying causal subgraph $G_C^*$, then

Figures (7)

  • Figure 1: Same substructure in different brain regions may reflect distinct functional implications.
  • Figure 2: The framework of BrainOOD.
  • Figure 3: Edge score map visualization for ID/OOD checkpoints on ID/OOD test set of ABIDE dataset. VIS = visual network; SMN = somatomotor network; DAN = dorsal attention network; VAN = ventral attention network; LN = limbic network; FPCN = frontoparietal control network; DMN = default mode network.
  • Figure 4: The visualization of the top 10 connections with the highest score on ABIDE OOD set.
  • Figure 5: Comparison with graph OOD methods in terms of test OOD accuracy across 10 folds on ABIDE and ADNI datasets.
  • ...and 2 more figures

Theorems & Definitions (3)

  • Theorem 4.1
  • Theorem B.1: Restatement of Theorem \ref{['thm:gib_loss']}
  • proof