PaperlayerView exact revision on arXiv

Source-linked AI summary

Graph Neural Network for Interpreting Task-fMRI Biomarkers

Xiaoxiao Li, Nicha C. Dvornek, Yuan Zhou, Juntang Zhuang, Pamela Ventola, James S. Duncan

arXiv:1907.01661v2cs.LGcs.CVeess.IVstat.ML

TL;DR

Interpreting GNN features in fMRI brain graphs is important for ASD biomarker discovery, but existing graph approaches lacked natural generalization to unseen nodes and graph-specific explanation tools. The paper proposes an inductive GNN pipeline for classifying task-fMRI graphs and identifying important regions and subgraphs without resampling occluded features. It reports improved accuracy and interpretability over a baseline, with robustness across atlases and hyperparameters.

  • Problem

    Existing graph approaches lack natural generalization to unseen nodes, and graph-specific tools for interpreting GNNs are described as unavailable.

    Source: §1 · p. 2

  • Method

    An inductive GNN classifies task-fMRI graphs, then scores subgraphs and ROIs from model responses without resampling occluded features.

    Source: §1 · p. 2

  • Results

    The framework achieved improved accuracy and more interpretable features than the baseline method, and performed robustly across atlases and hyperparameters.

    Source: §4 · p. 8

  • Takeaways & Limitations

    Orbitofrontal and ventromedial prefrontal cortices recur across tested atlases and cluster counts and are associated with social motivation.

    Source: §3.4 · p. 7

Abstract

from arXiv · show

Finding the biomarkers associated with ASD is helpful for understanding the underlying roots of the disorder and can lead to earlier diagnosis and more targeted treatment. A promising approach to identify biomarkers is using Graph Neural Networks (GNNs), which can be used to analyze graph structured data, i.e. brain networks constructed by fMRI. One way to interpret important features is through looking at how the classification probability changes if the features are occluded or replaced. The major limitation of this approach is that replacing values may change the distribution of the data and lead to serious errors. Therefore, we develop a 2-stage pipeline to eliminate the need to replace features for reliable biomarker interpretation. Specifically, we propose an inductive GNN to embed the graphs containing different properties of task-fMRI for identifying ASD and then discover the brain regions/sub-graphs used as evidence for the GNN classifier. We first show GNN can achieve high accuracy in identifying ASD. Next, we calculate the feature importance scores using GNN and compare the interpretation ability with Random Forest. Finally, we run with different atlases and parameters, proving the robustness of the proposed method. The detected biomarkers reveal their association with social behaviors. We also show the potential of discovering new informative biomarkers. Our pipeline can be generalized to other graph feature importance interpretation problems.

1 Introduction

The paper addresses the need to identify ASD biomarkers and the lack of generalization to unseen nodes and graph-specific explanation tools in existing approaches. It proposes an inductive GNN pipeline that classifies task-fMRI graphs and identifies important subgraphs and regions without resampling occluded features.

1 Introduction

  • Existing graph approaches lack natural generalization to unseen nodes, and available CNN explanation algorithms do not directly explain graph inputs.The paper also notes that no tool then existed to interpret and explain GNNs.Source: §1 · p. 2
  • The proposed method integrates connectivity, geometric, anatomical, and task-fMRI information into graphs, avoiding the need to predetermine the best features and connectivity measures.The authors describe those choices as ambiguous given task-fMRI’s complex structure.Source: §1 · p. 2
  • Its inductive GNN can classify whole graphs and be validated on subgraphs, scoring subgraphs and ROIs from model responses without resampling occluded features.The paper defines important subgraphs and ROIs as biomarkers in this setting.Source: §1 · p. 2

2 Methodology

The method represents task-fMRI brain networks as attributed multigraphs and uses a GNN that inductively learns node representations from neighboring nodes. The graph inputs combine regional measurements with functional and geometric edge information.

2.1 Graph Definition

  • Each brain ROI, parcellated from T1 structural MRI, becomes a node in an undirected multigraph with node and edge attributes.Node attributes include connectivity degree, GLM coefficients, task-fMRI mean and standard deviation, and ROI-center coordinates.Source: §2.1 · pp. 2–3
  • Edges encode Pearson and partial correlations plus a geometric-distance measure, with partial-correlation values thresholded at the 95th percentile for sparsity.Node features are Box-Cox transformed using parameters learned on training data and applied to training and testing sets.Source: §2.1 · p. 3

2.2 Graph Neural Network (GNN) Classifier

  • The GNN inductively learns node representations by recursively aggregating and transforming neighboring-node feature vectors.The architecture specifies node attributes, edge attributes, and kernel sizes.Source: §2.2 · p. 3

Convolutional Layer

The convolutional layer builds node embeddings by aggregating neighborhood information, using node degree and edge attributes in the message-passing process.

Convolutional Layer

  • The convolution follows Message Passing Neural Networks, uses node degree in the embedding, and is invariant to graph symmetries.The embedded representation corresponds to the lth convolutional layer.Source: Convolutional Layer · p. 3
  • For each node, the layer aggregates information from its one-hop neighborhood and applies a nonlinear activation, specified as ReLU.The passage identifies the neighborhood and activation used in the layer definition.Source: Convolutional Layer · p. 3
  • An MLP maps edge attributes to a matrix used in the message transformation.The mapped matrix has dimensions d(l) × d(l−1).Source: Convolutional Layer · p. 3

Pooling Aggregation Layer

The pooling layer reduces graph size by ranking nodes with a learnable projection and retaining the highest-scoring nodes. Its operations preserve sparsity through feature gating and slicing.

Pooling Aggregation Layer

  • A learnable vector projects node attributes to scores, with lower-scoring nodes receiving less feature retention.The operation maps an input graph to a pooled graph.Source: Pooling Aggregation Layer · p. 4
  • The layer selects the top k scores and gates the retained node features with tanh while leaving the edge matrix unchanged.The selected indices determine which node features are retained.Source: Pooling Aggregation Layer · p. 4
  • Projection, pointwise multiplication, and slicing preserve sparsity; the pooling vector is constrained to unit L2 norm through an additional regularization loss.The norm constraint is described as a difference from the approach in [2].Source: Pooling Aggregation Layer · p. 4

Readout Layer

The readout converts each convolution-pool block’s output graph into a fixed-size summary, then combines summaries for prediction.

Readout Layer

  • Each convolution-pool block’s output graph is summarized in a fixed-size representation.This provides a flattening operation for the graph output.Source: Readout Layer · p. 4
  • The block summary uses elementwise maximum over node feature vectors and concatenates the result with another node-feature summary.The passages identify the node count and feature vectors, but do not specify the other summary’s operation.Source: Readout Layer · p. 4
  • Summaries from all blocks are concatenated and submitted to an MLP for final predictions.Source: Readout Layer · p. 4

2.3 Explain Input Data Sensitivity

The sensitivity-explanation procedure clusters the whole-brain graph into subgraphs, evaluates their predictive power, and assigns importance scores to ROIs.

2.3 Explain Input Data Sensitivity

  • The procedure first clusters the whole-brain graph into subgraphs.Source: §2.3 · p. 4
  • It investigates the predictive power of each subgraph.Source: §2.3 · p. 4
  • It then assigns an importance score to each ROI.Source: §2.3 · p. 4

Network Community Clustering

The method stacks graph connectivity matrices across instances and applies non-negative PARAFAC decomposition to identify overlapping functional brain networks and region memberships.

Network Community Clustering

Graph Salience Mapping

After decomposing brain networks into community subgraphs, the method scores each community using its predictive evidence for the correct class, then aggregates scores to estimate node importance.

Graph Salience Mapping

  • Brain networks are decomposed into community subgraphs, which receive importance scores through salience mapping.Source: Graph Salience Mapping · p. 5
  • Each community is evaluated by inputting its subgraph directly into the pretrained classifier to estimate class probabilities.The classifier’s predicted probability for the original graph provides the comparison context.Source: Graph Salience Mapping · p. 5
  • Positive ECC supports the classifier’s correct-class prediction, while negative ECC counts against it; node scores aggregate ECC across communities containing that node.A Laplace correction is used to avoid zero denominators, and tanh bounds ECC.Source: Graph Salience Mapping · p. 5

3 Experiments and Results

The experiments evaluate task-fMRI graph classification and biomarker interpretation, then test whether the identified biomarkers persist across atlases and cluster counts.

3.1 Data Acquisition and Preprocessing

  • The study used task-fMRI from 75 children with ASD and 43 age- and IQ-matched healthy controls, augmenting the data into 750 ASD and 860 HC graphs.Five-fold splits were made by subject; each node had 10 attributes and each edge had 3.Source: §3.1 · p. 5

3.2 Step 1: Train ASD/HC Classification Model

  • 83.6% GNN versus 77.4% RF testing accuracy; the proposed model significantly outperformed alternative methods.The authors attribute the performance advantage to embedding high-dimensional features through structural relationships.Source: Table 1 · p. 6, §3.3 · p. 6, §3.2 · p. 6

3.3 Step 2: Interpret and Explain Biomarkers

  • RF and GNN shared nine of their top 30 ROIs; GNN assigned more uniform node-attribute importance, with biological-motion β1 ranked highest.The reported attribute scores also align similar task-fMRI mean/standard-deviation features and coordinates.Source: Figure 3 · p. 6, §3.3 · p. 7
  • The two highest-scoring subgraphs were associated with high-level social behaviors, and a model using their 19 ROIs achieved 78.9% test accuracy.That accuracy was comparable to using the whole graph in the best-model split.Source: §3.3 · p. 7, Figure 4 · p. 7

3.4 Evaluation: Robustness Discussion

  • Across the tested atlases and cluster counts, the orbitofrontal and ventromedial prefrontal cortices were selected in every case.The authors describe these regions as related to social motivation and previously associated with ASD.Source: §3.4 · p. 7, Figure 5 · p. 8

4 Conclusion and Future Work

The paper concludes that its task-fMRI GNN framework improves accuracy and interpretability over the baseline, with robust results across atlases and hyperparameters.

4 Conclusion and Future Work

  • The framework showed robust performance across atlases and hyperparameters, while future work will test additional hyperparameters, functional atlases, and graph definitions.The authors also identify resting-fMRI biomarker discovery and vessel cancer detection as possible applications.Source: §4 · p. 8
  • The authors conclude that their framework achieved improved accuracy and more interpretable features than the baseline method.Source: §4 · p. 8
  • The pipeline is presented as generalizable to other feature-importance analysis problems, including resting-fMRI biomarker discovery and vessel cancer detection.Source: §4 · p. 8

Related papers

Loading 1907.01661v2…