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
Irregular medical patient graphs vary in neighborhood density and local structure, and constant-size filters might not produce comparable features. InceptionGCN combines spectral filters at multiple neighborhood sizes; it improved over baseline by an average margin of 4.12% on TADPOLE, while results were comparable and lower on ABIDE.
Problem
Irregular patient graphs vary in neighborhood density and local structure, so constant-size filters may not produce comparable features across the graph.
Source: §1 · p. 2Method
InceptionGCN uses spectral convolution filters with different neighborhood sizes, merged within inception modules to capture local and broader context in heterogeneous graphs.
Source: §1 · p. 2, §2.3 · p. 5Results
InceptionGCN improved over baselines by an average margin of 4.12% on TADPOLE, while both methods had comparable, considerably lower performance on ABIDE.
Source: §3.2 · p. 7, §4 · p. 9Takeaways & Limitations
The findings support multi-scale feature learning for disease classification, with gains dependent on dataset in these evaluations.
Source: §4 · pp. 8–9
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 by using spectral convolutions with multiple kernel sizes to learn cluster-specific features. The paper studies Alzheimer’s and Autism prediction from complementary imaging and non-imaging data and reports improved performance and convergence over regular GCN architectures.
Introduction
- The study targets Alzheimer’s and Autism disease prediction using complementary imaging and non-imaging data, framing prediction as node-level classification on a patient-similarity graph.Patients are nodes, similarity weights define edges, and imaging modalities provide graph features.Source: §1 · pp. 1–2
- Because patient graphs have varying neighbor counts and local density, the authors propose InceptionGCN, which applies spectral convolutions with multiple kernel sizes to learn cluster-specific features.The model seeks similar features within clusters and different features between clusters.Source: §1 · p. 2
- Prior approaches use constant-size filters [6], receptive paths [7], layer concatenation [8], or fixed, hand-designed, or aggregator-based features, leaving graph heterogeneity insufficiently addressed.The paper also notes that the approach in requires a pre-defined node order that is difficult to obtain.Source: §1 · p. 2
- InceptionGCN is reported to improve performance and convergence over regular GCNs, with validation on artificial and clinical data and robustness across graph-adjacency construction approaches.The authors analyze how graph structure and filter size interact on one artificial and two public medical datasets.Source: §1 · pp. 2–3
2 Methodology
InceptionGCN builds patient affinity graphs and applies spectral filters with varying receptive fields, combining parallel outputs to represent neighborhood structure at multiple scales.
2 Methodology
- Unlike traditional fixed-filter models, InceptionGCN varies filter sizes across graph-convolution layers to address classes with distinct variance or heavy overlap.The approach targets disease classification using varied associations from different graphs in the population.Source: §2 · p. 3
2.1 Affinity graph construction
- The affinity graph spans training and testing patients, uses non-imaging data to define binary connections, and weights those edges by imaging-feature correlation distance.The final affinity matrix is the Hadamard product of the weight and edge matrices.Source: §2.1 · pp. 3–4
2.2 Mathematical background of spectral convolution and localization of filters for inception modules
- Chebyshev-polynomial spectral filters avoid costly Fourier transforms, and filters represented by degree-k Laplacian polynomials are exactly k-hop localized [6,1] [12].Spectral convolution is defined in the graph Fourier domain using the graph Laplacian’s eigenvectors and eigenvalues [6].Source: §2.2 · p. 4
2.3 Inception modules
- Each inception module applies S filters with varying neighborhood sizes and merges their outputs by concatenation or max-pooling.The architecture stacks M modules with Sm parallel GC-layers per module and ReLU; cross-entropy on labeled nodes trains through graph signal diffusion to unlabeled nodes.Source: §2.3 · p. 5
3 Experiments and Results
Experiments examine kernel-size sensitivity and compare InceptionGCN with baselines on two medical datasets and simulated data, finding dataset- and condition-dependent results.
3.1 Datasets
- TADPOLE includes 557 patients with 354 multimodal features across three diagnostic classes, while ABIDE uses 871 subjects across normal and ASD classes.TADPOLE features span imaging, cognitive, CSF, clinical, demographic, and APOE data; ABIDE aggregates data from 20 sites and follows baseline preprocessing.Source: §3.1 · pp. 5–6
3.2 Experiments on medical datasets
- Performance changes by an average margin of 8% across kernel-size choices, and different affinity graphs yield varied results for the same kernel combination.Accuracy increases with k before becoming consistent; for most graphs, k1 > k2 performs best, with the initial layer filtering global features.Source: §3.2 · p. 6, Figure 2 · p. 7
- InceptionGCN outperforms baseline methods by an average margin of 4.12% on TADPOLE, but is comparable to and does not outperform on ABIDE.On ABIDE, the mixed graph with feature-based edge weighting performs worse than the weighting case; the paper attributes the challenge to images collected across different sites.Source: §3.2 · p. 7
3.3 Experiments on simulated data
- InceptionGCN outperforms the baseline in most simulated settings, with superiority mainly when both classes have high variance; random features drastically reduce performance for all models.For clearly separable clusters, the baseline is less sensitive to kernel size; with higher variance it becomes more sensitive, and larger receptive fields can produce generalized global features.Source: §3.3 · p. 8
4 Discussion and Conclusion
InceptionGCN uses multiple kernel sizes to capture heterogeneous graph context and support multi-scale feature learning. It achieved a clear margin over baselines on TADPOLE, while its ABIDE performance was comparable to the baseline and lower overall.
Architecture contribution
- Multiple kernel sizes let InceptionGCN capture local and global heterogeneous graph context, improving multi-scale feature learning from shared input features and graphs.The authors report that heterogeneous features also help the model learn better classification filters.Source: §4 · pp. 8–9
Dataset results
- On TADPOLE, InceptionGCN had a clear margin over baselines; on ABIDE, both methods performed comparably, with considerably lower accuracy than on TADPOLE.The comparison covers Alzheimer’s and Autism disease prediction on two publicly available medical datasets.Source: Table 2 · p. 9, §4 · p. 9, Table 3 · p. 10
Simulated data
- In simulated data, InceptionGCN was superior with large variances and overlapping class clusters, while random node features sharply reduced performance and impaired convergence.Both GCN models were sensitive to data variance; the simulations also tested cases where graph connectivity alone carried class information.Source: §4 · pp. 9, 11, Table 1 · p. 9
Influence of the graph
- On ABIDE, site-based affinity forms 20 disjoint clusters with limited clinical relevance, while averaging affinity graphs performed worst overall and may erase graph-specific semantics.The authors note that ranking affinity graphs during training is a more elegant combination approach [16].Source: §4 · p. 10
Influence of the features
- Feature quality affected performance: ABIDE features were less cluster-distinguishing than TADPOLE’s, so feature-similarity graph construction hurt ABIDE, while meaningful TADPOLE features boosted performance.Randomly sampled node features also caused a drastic performance drop and large standard deviation, indicating poor filter learning and non-convergence.Source: §4 · p. 11
Influence of the kernel size
- Kernel-size sensitivity depended on feature and graph conditions: larger k helped with high-variance classes or class-indiscriminative features and a less relevant graph, but mattered little for compact clusters.With class-separable features, larger k produced more compact features; InceptionGCN’s multiple k values captured discriminative features for classes with large, different variances.Source: §4 · p. 11
Sequential model vs. InceptionGCN
- With early stopping at window size 25, InceptionGCN converged 1.63 times faster in epochs than the baseline, and the aggregator function affected performance.The baseline setups enumerate the possible combinations of the proposed model’s final output.Source: §4 · p. 11
Future scope
- Future work includes inductive out-of-sample inference, integrating multiple affinity graphs, and optimizing the model’s preprocessing, hidden-unit counts, and number of inception modules.The proposed preprocessing improvement would learn neighborhood values k.Source: §4 · p. 11