[Paper Review] Graph Saliency Maps through Spectral Convolutional Networks: Application to Sex Classification with Brain Connectivity

Salim Arslan, Sofia Ira Ktena+2 more|arXiv (Cornell University)|Jun 5, 2018

Functional Brain Connectivity Studies16 references4 citations

TL;DR

This paper proposes a novel visual attribution method for graph convolutional networks (GCNs) using spectral convolutions and class activation mapping to identify salient brain regions in functional connectivity networks without node-level labels. Applied to a sex classification task on UK Biobank rs-fMRI data (n > 5,000), the method robustly highlights the default mode network—particularly nodes 21, 5, 13, and 7—as most predictive of sex, with high reproducibility across runs and strong neurobiological relevance.

ABSTRACT

Graph convolutional networks (GCNs) allow to apply traditional convolution operations in non-Euclidean domains, where data are commonly modelled as irregular graphs. Medical imaging and, in particular, neuroscience studies often rely on such graph representations, with brain connectivity networks being a characteristic example, while ultimately seeking the locus of phenotypic or disease-related differences in the brain. These regions of interest (ROIs) are, then, considered to be closely associated with function and/or behaviour. Driven by this, we explore GCNs for the task of ROI identification and propose a visual attribution method based on class activation mapping. By undertaking a sex classification task as proof of concept, we show that this method can be used to identify salient nodes (brain regions) without prior node labels. Based on experiments conducted on neuroimaging data of more than 5000 participants from UK Biobank, we demonstrate the robustness of the proposed method in highlighting reproducible regions across individuals. We further evaluate the neurobiological relevance of the identified regions based on evidence from large-scale UK Biobank studies.

Motivation & Objective

• To develop a visual attribution method for graph-structured data to identify brain regions critical for classification without node-level supervision.
• To demonstrate the method’s robustness in identifying consistent, reproducible brain regions across subjects and training runs.
• To validate the neurobiological relevance of identified regions using large-scale UK Biobank connectome data and prior sex-difference studies.
• To extend class activation mapping from Euclidean images to non-Euclidean graph data, specifically functional brain networks.
• To provide a generalizable framework for ROI identification in graph-centric deep learning tasks in neuroscience.

Proposed method

• The method employs spectral convolutional networks (GCNs) to process functional connectivity graphs derived from resting-state fMRI data.
• It applies class activation mapping (CAM) to the final convolutional layer features to attribute saliency to individual nodes (brain regions).
• Node saliency is computed based on weighted activations of feature maps, with higher weights indicating greater contribution to the classification decision.
• The approach is applied to a binary sex classification task, where each graph represents a subject’s functional connectivity network.
• Saliency scores are averaged across subjects and multiple training runs to assess consistency and reproducibility.
• Neurobiological relevance is evaluated by mapping salient nodes to resting-state networks (RSNs) and comparing with group-averaged connectomes from UK Biobank.

Experimental results

Research questions

• RQ1 Can class activation mapping be effectively adapted to graph-structured data to identify salient nodes in functional brain networks?
• RQ2 Are the salient brain regions identified by the method consistent across different subjects and training runs?
• RQ3 Do the identified salient nodes correspond to biologically meaningful brain networks, such as the default mode network?
• RQ4 Is the method robust to variations in training initialization and hyperparameters?
• RQ5 Can the method detect known sex differences in functional connectivity without prior labeling of nodes?

Key findings

• The method consistently identifies the same four brain regions—nodes 21, 5, 13, and 7—as most salient across all subjects and multiple training runs, with over 90% of runs ranking them in the top three.
• These regions are part of the default mode network, a known functional network associated with internal cognition and self-referential processing.
• The default mode network shows stronger functional connectivity in females, aligning with prior UK Biobank findings and supporting the neurobiological relevance of the identified regions.
• The saliency maps are highly reproducible across different random seeds and cross-validation folds, indicating robustness to training variation.
• The method achieves state-of-the-art accuracy in sex classification using functional connectivity graphs, validating its predictive power.
• The approach successfully identifies biologically meaningful ROIs without requiring node-level labels, demonstrating its utility for unsupervised ROI discovery in connectomics.