Human Brain Functional Modules: A Novel Detection Algorithm and A Multiresolution Comparative Study
Poster No:
2142
Submission Type:
Abstract Submission
Authors:
Xitian Chen1, Yining Zhang1, Zhengjia Dai1, Ying Lin1
Institutions:
1Sun Yat-sen University, Guangzhou, Guangdong
First Author:
Co-Author(s):
Introduction:
Combining resting-state functional MRI (R-fMRI) and graph theory, a functional brain network, where nodes represent brain regions and edges represent functional connectivity (FC), is organized into modules defined by dense FC within each module and sparse FC between them (Meunier et al., 2009). The resolution limit of community detection algorithms (Sporns, 2018) leads to a poor explanation of the complex brain functional modular structure (Wang et al., 2021; Lin et al., 2018). An avenue is to introduce a resolution-related parameter that produces multiple solutions spanning from coarser to finer resolutions. The developing computational intelligence algorithms paved the way for a resolution-parameter-free design, which decomposes the definition of modularity into two objective functions and processes a multiobjective search. Therefore, we developed a multiobjective evolutionary algorithm (MOEA) to detect functional modular structures and conducted a multiresolution comparison of algorithms and their cognitive predictions.
Methods:
Data were obtained from the HCP dataset, consisting of 310 healthy adults. After the HCP preprocessing pipeline, R-fMRI data were regressed out linear trends and nuisance signals and filtered (0.01~0.1 Hz). We utilized Power's atlas (Power et al., 2011) and calculated the sparse group-average no-negative matrix (sparsity = 25.2%) to detect modular structure. MOEA (Fig 1A) produced a Pareto front (PF) comprising non-dominated partitions that attained near-optimal yet distinct trade-offs between the conflicting Kernel Kmeans (KKM) and Ratio cut (RC) (Lin et al., 2018). For the multiresolution comparison of MOEA and ten classic algorithms (Fig 1B), partitions were organized into a three-level resolution hierarchy (low/medium/high) based on the maximum-information-gain criterion. The representative partition for each algorithm at each level was selected as the one having the highest average normalized mutual information with others. We adapted the connectome-based predictive modeling (Shen et al., 2017) into a module-based approach for cognitive predictions. Modular statistics (herein, intra-module strength z-score and participation coefficient) were used to predict the principal component scores of 58 behavioral measures.
Results:
We plotted the estimated 225 network partitions generated by the 11 algorithms in the RC-KKM space (Fig 1C). The PF ranges from the 'cohesiveness-preferred' upper left to the 'segregation-preferred' lower right. Only a few AFG's and RN's partitions were superior to the reference PF. Resolution levels L1~L3 of our hierarchy consisted of partitions with 2~6, 7~10, and 11~29 modules, respectively. While single-solution algorithms' partitions were all at L1, multi-solution algorithms except for RN could generate partitions across all levels but with different preferences (Fig 2A). Fig 2B shows the MOEA's representative partitions at each level. Two highlighted principal components, Comp1 and Comp2, were derived from behavioral data. While only RB's partitions could successfully predict Comp1 in a few repetitions, the prediction for Comp2 succeeded on most partitions (significant predicted-observed correlations) (Fig 2C). As the resolution increased, the prediction accuracy of Comp2 worsened in AFG, RB, and StrC, but improved again after a decline in SpeC and MOEA (Fig 2D). Interestingly, the most accurate prediction results were obtained by SpeC at L3 and MOEA at L1.
Conclusions:
This study constructs a multiresolution comparison of module detection algorithms. We use a reference front comprised of partitions generated by the proposed resolution-parameter-free multiobjective optimizer. By establishing a three-level multiresolution hierarchy and extracting representative partitions at different resolutions to predict task performance, we found the resolution preferences of multi-solution algorithms and revealed their task-performance prediction advantages.
Modeling and Analysis Methods:
Classification and Predictive Modeling 2
fMRI Connectivity and Network Modeling
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
Microcircuitry and Modules 1
Keywords:
FUNCTIONAL MRI
Machine Learning
NORMAL HUMAN
Other - human connectome, module, community detection
1|2Indicates the priority used for review
Provide references using author date format
Meunier, D., et al. (2009), ‘Hierarchical modularity in human brain functional networks’, Frontiers in neuroinformatics, vol. 3, no. 1, pp. 571.
Power, J.D., et al. (2011), ‘Functional Network Organization of the Human Brain’, Neuron, vol. 72, pp. 665–678.
Sporns, O. (2018), ‘Graph theory methods: applications in brain networks’, Dialogues in clinical neuroscience.
Wang, R., et al. (2021), ‘Segregation, integration, and balance of large-scale resting brain networks configure different cognitive abilities’, Proceedings of the National Academy of Sciences, vol. 118, no. 23, e2022288118.