Explainable Graph Neural Network on Imaging Genetics for Schizophrenia Classification
Poster No:
1434
Submission Type:
Abstract Submission
Authors:
Badhan Mazumder1, Vince Calhoun2, Dong Hye Ye1
Institutions:
1Georgia State University, ATLANTA, GA, 2GSU/GATech/Emory, Decatur, GA
First Author:
Co-Author(s):
Introduction:
Investigation into the brain's structure and functional connections could provide vital insights into neuropsychiatric disorders like schizophrenia (SZ). Genetic markers, such as single nucleotide polymorphisms (SNP), are known to play a significant role in identifying SZ patients due to their strong heritability. In this study an explainable graph neural network (GNN) framework is introduced to classify SZ patients. By obtaining rich data embeddings from multimodal graph-structured data formed with the subject-wise functional network connectivity (FNC) obtained from resting-state fMRI (rs-fMRI), structural connectivity (SC) derived from diffusion magnetic resonance imaging (dMRI), and SNPs. Experimental outcomes showed that in contrast to unimodal methods, our multimodal approach notably enhanced the classification of SZ patients against healthy controls (HC).
Methods:
Data preprocessing
Subject-specific FNC matrices were obtained via the Neuromark pipeline [1]. To derive SC matrices, we computed diffusion tensors with dtifit (FSL) [2], followed by deterministic tractography generation of the whole brain using track (CAMINO) [3]. To create a compatible atlas in local space, firstly we performed inverted spatial normalization to the NeuroMark atlas [1], then warped the native fractional anisotropy map onto the standard MNI space. Finally, we isolated the streams passing through each pair of atlas ROIs and counted them individually.
Proposed method
As depicted in Figure 1, we constructed graph structure data where the edges were formed by applying a k-nearest neighbors (KNN) graph on subject-specific SC to consider crucial connections only. Since a large fraction of SNP data is unrelated to SZ, we employed a subset of 4943 SNPs, that by the findings of the psychiatric genomic consortium have been observed for SZ risk [4]. A 1-D convolutional neural network (CNN) followed by layer-wise relevance propagation (LRP) [5] was applied to identify the top 100 SZ-linked SNPs. Then the corresponding FNC matrix was concatenated with the selected SNPs and employed as node features. After that, we input the multimodal graph data to our GNN framework that includes 5 graph convolution (GCN) [6] layers, a global mean pooling layer followed by a linear layer, and a softmax for classification. Additionally, to explain, we employed GNNExplainner [7] to identify the important edges of the graph which were highly significant for SZ classification.
Subject-specific FNC matrices were obtained via the Neuromark pipeline [1]. To derive SC matrices, we computed diffusion tensors with dtifit (FSL) [2], followed by deterministic tractography generation of the whole brain using track (CAMINO) [3]. To create a compatible atlas in local space, firstly we performed inverted spatial normalization to the NeuroMark atlas [1], then warped the native fractional anisotropy map onto the standard MNI space. Finally, we isolated the streams passing through each pair of atlas ROIs and counted them individually.
Proposed method
As depicted in Figure 1, we constructed graph structure data where the edges were formed by applying a k-nearest neighbors (KNN) graph on subject-specific SC to consider crucial connections only. Since a large fraction of SNP data is unrelated to SZ, we employed a subset of 4943 SNPs, that by the findings of the psychiatric genomic consortium have been observed for SZ risk [4]. A 1-D convolutional neural network (CNN) followed by layer-wise relevance propagation (LRP) [5] was applied to identify the top 100 SZ-linked SNPs. Then the corresponding FNC matrix was concatenated with the selected SNPs and employed as node features. After that, we input the multimodal graph data to our GNN framework that includes 5 graph convolution (GCN) [6] layers, a global mean pooling layer followed by a linear layer, and a softmax for classification. Additionally, to explain, we employed GNNExplainner [7] to identify the important edges of the graph which were highly significant for SZ classification.
·Overview of Proposed Method
Results:
For validation, we utilized a subset (Total:165; SZ:93, HC:72) of the Function Biomedical Informatics Research Network (FBIRN) [8] dataset with 80:20 training-testing split ratio.
We evaluated using 5-fold cross-validation against 3 different indices: accuracy, precision, and f1-score. For the k-NN graph on SC, different k values (3, 5, 10, and 20) were utilized. Among them, k=10 resulted in the highest performance. Results shown in Figure 2a indicate that compared with the unimodal (SC and FNC separately) model the multimodal (SC+FNC) model was giving higher performance which improved to 73.38% (accuracy), 75.10% (precision) and 72.79% (f1-score) when SNP was also considered.
As shown in Figure 2b, from our generated explanations for SZ patients, we found significance in default mode (DMN), visual (VSN), and cognitive-control (CON) network which is consistent with the clinical findings. We also conducted a group analysis using a 2-sample T-test (p<0.05) in-between generated explanations for HC and SZ and depicted the finding in Figure 2c which also indicates the same networks.
We evaluated using 5-fold cross-validation against 3 different indices: accuracy, precision, and f1-score. For the k-NN graph on SC, different k values (3, 5, 10, and 20) were utilized. Among them, k=10 resulted in the highest performance. Results shown in Figure 2a indicate that compared with the unimodal (SC and FNC separately) model the multimodal (SC+FNC) model was giving higher performance which improved to 73.38% (accuracy), 75.10% (precision) and 72.79% (f1-score) when SNP was also considered.
As shown in Figure 2b, from our generated explanations for SZ patients, we found significance in default mode (DMN), visual (VSN), and cognitive-control (CON) network which is consistent with the clinical findings. We also conducted a group analysis using a 2-sample T-test (p<0.05) in-between generated explanations for HC and SZ and depicted the finding in Figure 2c which also indicates the same networks.
·Classification Performance & Visualization of Important Connections
Conclusions:
We presented an explainable GNN framework that incorporates structural-functional connectomics and genetic information simultaneously, to improve SZ classification and interpret potential structural biomarkers. Obtained experimental findings showed improved performance gain in comparison to unimodal set-up which provides evidence of the robustness of our multimodal approach.
Disorders of the Nervous System:
Psychiatric (eg. Depression, Anxiety, Schizophrenia)
Genetics:
Genetics Other
Modeling and Analysis Methods:
Classification and Predictive Modeling 1
Connectivity (eg. functional, effective, structural) 2
Keywords:
Data analysis
FUNCTIONAL MRI
Machine Learning
Psychiatric Disorders
Schizophrenia
1|2Indicates the priority used for review
Provide references using author date format
2. Woolrich, Mark W., Saad Jbabdi, Brian Patenaude, Michael Chappell, Salima Makni, Timothy Behrens, Christian Beckmann, Mark Jenkinson, and Stephen M. Smith. "Bayesian analysis of neuroimaging data in FSL." Neuroimage 45, no. 1 (2009): S173-S186.
3. Alexander, D. C., K. Seunarine, S. Nedjati-Gilani, M. G. Hall, G. J. Parker, Y. Bai, and P. J. Cook. "Camino: Open-source diffusion-MRI reconstruction and processing." In Proceedings of the 14th Scientific Meeting of the International Society for Magnetic Resonance in Medicine (ISMRM), Seattle, WA, USA, pp. 6-12. 2006.
4. Yu, Hongkun, Thomas Florian, Vince Calhoun, and Dong Hye Ye. "Deep Learning From Imaging Genetics for Schizophrenia Classification." In 2022 IEEE International Conference on Image Processing (ICIP), pp. 3291-3295. IEEE, 2022.
5. Bach, Sebastian, Alexander Binder, Grégoire Montavon, Frederick Klauschen, Klaus-Robert Müller, and Wojciech Samek. "On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation." PloS one 10, no. 7 (2015): e0130140.
6. Kipf, Thomas N., and Max Welling. "Semi-supervised classification with graph convolutional networks." arXiv preprint arXiv:1609.02907 (2016).
7. Ying, Zhitao, Dylan Bourgeois, Jiaxuan You, Marinka Zitnik, and Jure Leskovec. "Gnnexplainer: Generating explanations for graph neural networks." Advances in neural information processing systems 32 (2019).
8. Keator, David B., Theo GM van Erp, Jessica A. Turner, Gary H. Glover, Bryon A. Mueller, Thomas T. Liu, James T. Voyvodic et al. "The function biomedical informatics research network data repository." Neuroimage 124 (2016): 1074-1079.