Clustintime: a python toolbox for spatio-temporal clustering of fMRI data
Poster No:
1757
Submission Type:
Abstract Submission
Authors:
Cristina Tobias Figuerola1, Eneko Uruñuela2, Fernando Pérez-Bueno3, Itxaso Aizpurua4, Laura De Frutos5, Lucia Manso5, Sylvia Yang5, Vicente Ferrer5, Cesar Caballero-Gaudes5
Institutions:
1Linköpings Universitet, Linköping, Östergötland, 2Basque Center on Cognition, Brain and Language, Donostia - San Sebastián, Gipuzkoa, 3Universidad de Granada, Granada, Spain, 4Independent researcher, Barcelona, Spain, 5Basque Center of Cognition, Brain and Language, San Sebastián, Spain
First Author:
Co-Author(s):
Introduction:
The identification of synchronous BOLD fMRI signals is usually done through spatial clustering, allowing researchers to identify functional brain networks or parcellations9. There is an increasing interest in employing clustering algorithms in the temporal domain to identify instances of BOLD activity sharing similar spatial patterns [2]. For that purpose, we created Clustintime in 2022 and present here its latest release [1].
Clustintime stands as an open-source toolboxto automatically identify spatio-temporal patterns of neuronal activity in fMRI data. With a GPL-3.0 license, coded in python3 and based on Scikit-learn [3], infomap [4], and networkx. Comprising 5 modules, it offers a range of tools for diverse time-clustering analyses, meeting various user needs. This release introduces the capability to cluster multiple datasets simultaneously, while offering the possibility of consensus clustering across all algorithms.
Clustintime stands as an open-source toolboxto automatically identify spatio-temporal patterns of neuronal activity in fMRI data. With a GPL-3.0 license, coded in python3 and based on Scikit-learn [3], infomap [4], and networkx. Comprising 5 modules, it offers a range of tools for diverse time-clustering analyses, meeting various user needs. This release introduces the capability to cluster multiple datasets simultaneously, while offering the possibility of consensus clustering across all algorithms.
Methods:
In this release, Clustintime maintains its user-friendly 3-step pipeline (matrix computation-clustering-visualization), now improved with an optional 4th feature for consensus clustering. Internally, all the modules have been modified to achieve a better performance.
The main module underwent significant changes, absorbing mandatory features previously within the processing module (i.e: choosing a positive, negative or whole-signal analysis anddistance matrix computation with or without the sliding window approach). Additionally, it now supports simultaneous loading of data from multiple files, enabling researchers to employ clustering methods for studying groups and/or multiple run analyses.
The processing module has evolved into an optional feature, providing users with an additional choice rather than a compulsory step. These updates transformed the toolbox into a more modular tool, with improved code readability and facilitated collaboration for new contributors.
Further, a newly integrated consensus module enriches user experience with more robust outcomes than a singular execution of an algorithm on the entire dataset, especially if said algorithm employs heuristic methods. If selected, the clustering algorithm will be applied to, at least, 100 random subsamples of the data until it converges into a common solution.
The main module underwent significant changes, absorbing mandatory features previously within the processing module (i.e: choosing a positive, negative or whole-signal analysis anddistance matrix computation with or without the sliding window approach). Additionally, it now supports simultaneous loading of data from multiple files, enabling researchers to employ clustering methods for studying groups and/or multiple run analyses.
The processing module has evolved into an optional feature, providing users with an additional choice rather than a compulsory step. These updates transformed the toolbox into a more modular tool, with improved code readability and facilitated collaboration for new contributors.
Further, a newly integrated consensus module enriches user experience with more robust outcomes than a singular execution of an algorithm on the entire dataset, especially if said algorithm employs heuristic methods. If selected, the clustering algorithm will be applied to, at least, 100 random subsamples of the data until it converges into a common solution.
Results:
As an example, we applied Clustintime to a motor task of a single subject (100206) from the Human Connectome Project (HCP) [6] using the Infomap algorithm. The images were already preprocessed using a standard HCP pipeline including realignment, coregistration, spatial normalization and smoothing and had been previously analyzed with multivariate sparse Paradigm Free Mapping using stability-selection [5].
Clustintime grouped the BOLD data volumes into multiple clusters (Figure 2A), automatically revealing activation maps (Figure 2B-E) associated with the tasks. Among the 13 clusters, a temporal pattern can be identified when observing the first 4 clusters. Figure 2B illustrates cluster 1, displaying activation in the region responsible for left foot movement. Figure 2C depicts cluster 2 with activation on the regions responsible for finger tapping. Figure 2D depicts cluster 3 showing activation inregions associated with tongue movement. Figure 2E portrays cluster 4 indicating activation in regions linked to left finger-tapping.
Clustintime grouped the BOLD data volumes into multiple clusters (Figure 2A), automatically revealing activation maps (Figure 2B-E) associated with the tasks. Among the 13 clusters, a temporal pattern can be identified when observing the first 4 clusters. Figure 2B illustrates cluster 1, displaying activation in the region responsible for left foot movement. Figure 2C depicts cluster 2 with activation on the regions responsible for finger tapping. Figure 2D depicts cluster 3 showing activation inregions associated with tongue movement. Figure 2E portrays cluster 4 indicating activation in regions linked to left finger-tapping.
Conclusions:
Clustintime, an open-source tool, empowers researchers to seamlessly integrate diverse tools for temporal clustering analyses of fMRI data. Its latest release enables the performance of these analyses across multiple fMRI volumes and introduces a consensus clustering feature. This toolbox gathers processing, clustering, and visualization methods to help researchers exploring events shaping the dynamics of brain function or identifying instances of artifact data. Backed by an open and expanding community, this project embraces contributions of any kind.
Modeling and Analysis Methods:
fMRI Connectivity and Network Modeling 1
Methods Development 2
Neuroinformatics and Data Sharing:
Informatics Other
Keywords:
FUNCTIONAL MRI
Informatics
Machine Learning
Open-Source Software
1|2Indicates the priority used for review
Provide references using author date format
2. Jo, Y. (2021). ‘The diversity and multiplexity of edge communities within and between brain systems’. Cell reports, 37(7), 110032. https://doi.org/10.1016/j.celrep.2021.110032
3. Pedregosa, F (2011) ‘Scikit-learn: Machine learning in Python’, the Journal of machine Learning research, 12, 2825-2830.
4. Thirion, B. (2014), ‘Which fMRI clustering gives good brain parcellations?’, Front. Neurosci. 8:167.
5. Uruñuela, E., Gonzalez-Castillo, J., Zheng, C., Bandettini, P., & Caballero-Gaudes, C. (2023). Whole-brain multivariate hemodynamic deconvolution for functional MRI with stability selection. Medical Image Analysis, 103010.
6. Van Essen, D. C (2013), ‘The WU-minn human connectome project: An overview.’, NeuroImage, 80 , 62–79. doi:10.1016/j.neuroimage.2013.05.041.