Source-linked AI summary
InceptionGCN: Receptive Field Aware Graph Convolutional Network for Disease Prediction
Anees Kazi, Shayan shekarforoush, S. Arvind krishna, Hendrik Burwinkel, Gerome Vivar, Karsten Kortuem, Seyed-Ahmad Ahmadi, Shadi Albarqouni, Nassir Navab
TL;DR
Disease prediction with multimodal medical data is complicated by heterogeneous graph structure, which fixed-size graph filters may not capture well. The paper proposes InceptionGCN, using spectral convolutions with multiple kernel sizes and an aggregation function, and evaluates it on clinical, artificial, and simulated settings. InceptionGCN shows a clear advantage on TADPOLE, while performance is comparable to the other method on ABIDE and depends on graph construction and feature informativeness.
Problem
Heterogeneous patient graphs contain varying neighborhood structures, so constant-size filters may not produce comparable semantic features for disease prediction.
Method
InceptionGCN applies multiple spectral filters with different neighborhood sizes to the same graph signals and aggregates their outputs.
Results
InceptionGCN has a clear performance margin over baselines on TADPOLE, while ABIDE shows comparable but considerably lower performance for both methods.
Takeaways & Limitations
Using different filter sizes helps learn local and global context in heterogeneous graphs, while meaningful graph affinities and features remain important for classification performance.
Abstract
from arXiv · showhide
Geometric deep learning provides a principled and versatile manner for the integration of imaging and non-imaging modalities in the medical domain. Graph Convolutional Networks (GCNs) in particular have been explored on a wide variety of problems such as disease prediction, segmentation, and matrix completion by leveraging large, multimodal datasets. In this paper, we introduce a new spectral domain architecture for deep learning on graphs for disease prediction. The novelty lies in defining geometric 'inception modules' which are capable of capturing intra- and inter-graph structural heterogeneity during convolutions. We design filters with different kernel sizes to build our architecture. We show our disease prediction results on two publicly available datasets. Further, we provide insights on the behaviour of regular GCNs and our proposed model under varying input scenarios on simulated data.
1 Introduction
InceptionGCN addresses heterogeneous patient graphs, where fixed-size GCN filters may not produce comparable semantic features across clusters. It introduces multi-sized spectral filters to capture local and broader graph structure for disease prediction.
- Application: The broader task is Alzheimer and Autism disease prediction using complementary imaging and non-imaging multimodal data.The framework uses patient relationships represented in a neighborhood graph for node-level classification.
- Motivation: Patient graphs vary in neighborhood count, density, and local topology, making constant-size filters potentially unsuitable for comparable feature learning.The paper frames this challenge through irregular graph structure and heterogeneous patient clusters.
- Proposed approach: InceptionGCN uses multiple kernel sizes on the same input to learn cluster-specific features across heterogeneous graph structures.Its filters target similar intra-cluster features and different inter-cluster features.
- Related work: The paper investigates receptive-field heterogeneity because earlier constant-filter GCNs do not address variation in graph structure.Related work includes selective receptive paths but provides limited focus on GCN receptive fields.
- Contributions: InceptionGCN is evaluated on artificial and clinical data, with reported improvements over regular GCN architectures in performance and convergence.The study also examines robustness to different graph-adjacency construction approaches.
- Proposed approach: The architecture processes feature matrix X with several graph-convolution layers using neighborhoods k1 through kS, then aggregates their outputs.This setup is illustrated for the TADPOLE affinity graph and the inception module.
2 Methodology
The methodology constructs patient affinity graphs from non-imaging similarities and imaging features, then applies localized spectral filters. Inception modules combine filters with varying receptive-field sizes to capture neighborhoods at different scales.
- Affinity graph construction: Patients are modeled as graph nodes, with non-imaging similarities defining edges and imaging modalities providing node features for node-level disease classification.
- Affinity graph construction: The affinity graph combines a binarized edge graph with feature-based edge weights through a Hadamard product.The edge graph encodes non-imaging similarity thresholds, while weights use correlation distance between node features.
- Spectral convolution: Spectral convolution applies learnable filters to graph signals through the graph Laplacian and Chebyshev polynomial parameterization.The polynomial form avoids the computationally prohibitive Fourier-transform matrix multiplication.
- Filter localization: A kth-order polynomial filter is exactly k-hop localized, incorporating neighbors within a specified graph distance.Filter localization is defined by considering all neighbors at distance k hops during spectral convolution.
- Inception modules: Inception modules use multiple filters with varying neighborhood sizes to combine close-proximity signals with broader neighborhood structure.Each filter has its own learnable parameters, and outputs are merged by concatenation or max-pooling.
3 Experiments and Results
Experiments on TADPOLE, ABIDE, and simulated data examine how graph construction, kernel sizes, and cluster variance affect spectral GCN performance. InceptionGCN improves over regular GCN on TADPOLE, performs comparably on ABIDE, and is especially beneficial in challenging simulated settings.
- Experimental setup: The study evaluates kernel-size sensitivity and InceptionGCN against baseline GCNs on two multimodal medical datasets and simulated data.The experiments also analyze graph-specific design choices and input scenarios.
- Datasets: TADPOLE contains 557 patients with 354 multimodal features for three-way classification of cognitively normal, mild cognitive impairment, and Alzheimer’s disease.Features include imaging, cognitive, cerebrospinal-fluid, clinical, demographic, and APOE information.
- Medical datasets: An 8% performance margin appears across TADPOLE combinations of k1 and k2, where smaller kernels capture local features and larger kernels capture global features.The best kernel pair for each graph is selected from k1 and k2 values ranging from 1 to 6 before constructing InceptionGCN.
- Medical datasets: InceptionGCN outperforms baseline GCNs on TADPOLE by an average margin of 4.12%.The comparison evaluates inception layers against baselines using repeated kernel sizes while accounting for graph-specific affinity.
- Medical datasets: On ABIDE, InceptionGCN performs comparably to the baseline but does not outperform it, while feature-based edge weighting performs worse in the mixed graph.The paper attributes this pattern to non-discriminative features and images collected across different sites.
- Simulated data: In simulated data, baseline sensitivity to kernel size increases with cluster variance, while larger receptive fields produce generalized global features and InceptionGCN is strongest in challenging settings.Randomly sampled node features cause drastic performance drops, large standard deviations, and failure to converge.
4 Discussion and Conclusion
InceptionGCN uses multiple kernel sizes to capture local and global context in heterogeneous graphs. Across evaluations, it improved TADPOLE performance over baselines, while ABIDE performance remained lower and comparable between methods.
- Model and findings: Multiple kernel sizes let InceptionGCN capture local and global context in heterogeneous graph structures.The architecture applies differently sized filters to the same input features and graph for multi-scale feature learning.
- Simulation analysis: Table 1 compares accuracy across simulated Gaussian classes by varying class variances, distinguishing settings where features and graph structure are correlated from graph-only settings.The table evaluates model behavior under different feature and connectivity relationships.
- Experimental evaluation: Table 2 reports mean stratified k-fold cross-validation accuracies for TADPOLE experiments and identifies the selected kernel pairs for each graph.The chosen [k1, k2] values are highlighted in the corresponding figure.
- Model and findings: InceptionGCN achieved a clear performance margin over baselines on TADPOLE, whereas ABIDE performance was comparable between methods and considerably lower.The paper attributes the differing outcomes to dataset and graph characteristics discussed for the two medical datasets.
- Graph influence: ABIDE’s site-based graph forms 20 disjoint clusters, limiting feature propagation to patients from the same site and reducing clinical relevance.The dataset combines images from 20 sites and imaging conditions, creating substantial heterogeneity.
- Experimental evaluation: Table 3 reports mean stratified k-fold cross-validation accuracy for ABIDE across graph setups, with baseline kernel pairs varying by affinity graph.The listed baseline pairs are [4,5], [6,5], and [4,4] for the specified graph configurations.
- Feature influence: Random features caused drastic performance drops and large standard deviations, while meaningful TADPOLE features improved graph generation and model performance.The results indicate that feature variance and informativeness affect convergence and classification performance.