Graph Saliency Maps through Spectral Convolutional Networks: Application to Sex Classification with Brain Connectivity (1806.01764v1)

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.

Graph Saliency Maps through Spectral Convolutional Networks: An Application in Brain Connectivity

This paper presents a methodology for identifying salient nodes in graph-structured data, specifically applied to brain connectivity networks. The authors focus on the utility of Graph Convolutional Networks (GCNs) enhanced with a visual attribution mechanism to achieve this task in a neuroscience context. By leveraging spectral convolutional networks, they attempt to map distinctive brain regions associated with sex classification based on functional connectivity networks, employing data from a significant cohort of over 5000 participants from the UK Biobank.

Methodology Overview

The paper outlines a novel approach combining spectral convolutional networks and class activation mapping (CAM) to discern key brain regions associated with sex-related differences in brain connectivity. The GCNs are structured to handle non-Euclidean data inherent to graph representations of brain networks. This technique enables the pinpointing of regions of interest (ROIs) on the connectivity graph of each subject, which significantly contributes to the prediction—achieving classification without pre-assigned node labels.

The spectral graph convolutions are conducted in the graph's Fourier domain, leveraging Chebyshev polynomial approximations for localized filtering, following Defferrard et al.'s methodology. This allows for efficient node feature processing and subsequent node classification by passing through several graph convolutional layers.

Class activation mapping is integrated into the GCN to generate saliency maps. The method effectively reveals which nodes (brain regions) the network deems important for classifying the sex of the individuals. This combines feature map activations from the final convolutional layer to produce node saliency scores. The approach identifies how specific nodes contribute to the classification decision, enabling a deeper understanding of functional connectivity differences.

Results and Implications

The paper reports an average classification accuracy of 88.06% across multiple runs with a negligible standard deviation, indicating robust performance. Notably, the network consistently highlighted regions within the default mode network as salient, corroborating existing literature on sex differences in functional brain connectivity. The identified regions displayed functional significance according to large-scale studies using UK Biobank data, thus reinforcing the method's neurobiological relevance.

From a methodological standpoint, this approach opens avenues for the deployment of GCNs in neuroscience, offering insights into the functional architecture of the brain under varying conditions and traits. Theoretical implications include advancing understanding of graph-centric neural networks, particularly in the field of human connectomics.

Future Directions

While the method provides a promising tool for understanding brain connectivity, its applicability extends beyond this initial domain. Future investigations could explore dynamic graph structures and the influence of other factors, such as age-related changes, potentially broadening its scope to additional neurological and psychological phenomena. Furthermore, validating this method in different settings and datasets will be crucial for establishing broader applicability and robustness.

In conclusion, this paper outlines a substantial step forward in applying graph neural network methodologies for neuroscientific exploration, with significant potential for far-reaching impacts in understanding complex neural processes.