Deep interpretable fMRI dynamics model for brain connectivity estimation
Poster No:
1411
Submission Type:
Abstract Submission
Authors:
Pavel Popov1, Giorgio Dolci2, Cristian Morasso2, Ilaria Boscolo Galazzo2, Gloria Menegaz2, Godfrey Pearlson3, David Danks4, Vince Calhoun5, Sergey Plis1
Institutions:
1Georgia State University, Atlanta, GA, 2University of Verona, Verona, Veneto, 3Yale University, New Haven, CT, 4University of California, San Diego, CA, 5GSU/GATech/Emory, Atlanta, GA
First Author:
Co-Author(s):
Introduction:
The complexity of neuroimaging data has driven the development of novel methods for its analysis with the goal of advancing our understanding of brain function and identifying biomarkers for neurological and psychiatric disorders. Machine learning (ML), as a robust tool for data analysis, has been extensively used in fMRI analysis, where it was proven to be useful in diagnostics of conditions such as autism, depression, schizophrenia, and Alzheimer's disease (Sharma, 2024), as well as demographic predictions like age and sex (Yeung, 2023). Despite this success, ML methods are often challenging to interpret when it comes to finding explanations for their decisions in the data. This challenge stems from the inherent complexity of the models, combined with the high dimensionality and low signal-to-noise ratio characteristic of fMRI data (Lindquist, 2008).
In this work, we present BrainDynaMo, a novel deep learning (DL) architecture for fMRI analysis with an emphasis on interpretability and explainability. Our architecture learns functional network connectivity (FNC) as part of its strategy to model brain dynamics during training. Unlike statistically-derived FNCs, these representations are learned with a classification task in mind, providing a more precise and task-relevant understanding of brain connectivity. Moreover, this approach allows to preserve more information on temporal dynamics inherent in fMRI time series, which is otherwise reduced to simple correlations.
In this work, we present BrainDynaMo, a novel deep learning (DL) architecture for fMRI analysis with an emphasis on interpretability and explainability. Our architecture learns functional network connectivity (FNC) as part of its strategy to model brain dynamics during training. Unlike statistically-derived FNCs, these representations are learned with a classification task in mind, providing a more precise and task-relevant understanding of brain connectivity. Moreover, this approach allows to preserve more information on temporal dynamics inherent in fMRI time series, which is otherwise reduced to simple correlations.
Methods:
A schematic overview of the BrainDynaMo model is shown in Figure 1. The main block of the model can be viewed as a variation of a recurrent neural network that keeps a distinct hidden state for each input channel and mixes them on updates using a learnable mixing matrix, which can be extracted along with the hidden state at each time point. This learning strategy allows us to model the information exchange between different brain regions with the mixing matrices, which effectively capture FNC patterns. We used these mixing matrices for classification and, to additionally ensure that the model learns the meaningful dynamic patterns in the data, we gave the model a task to predict the next input fMRI time point. Such design also allows us to pretrain the model solely for the input prediction task on a larger general population datasets.
In our experiments we trained BrainDynaMo for classification and input prediction tasks, and then used the learned FNC matrices in group analysis. We used the fMRI images from FBIRN (Function Biomedical Informatics Research Network) (Keator, 2016), COBRE (Center of Biomedical Research Excellence) (Çetin, 2014), BSNIP (Bipolar and Schizophrenia Network for Intermediate Phenotypes) (Tamminga, 2014), ABIDE (Autism Brain Imaging Data Exchange, release 1.0) (Di Martino, 2014), OASIS (Open Access Series of Imaging Studies, release 3.0) (Rubin, 1998), and HCP (Human Connectome Project, 1200 subjects release) (Van Essen, 2013).
In our experiments we trained BrainDynaMo for classification and input prediction tasks, and then used the learned FNC matrices in group analysis. We used the fMRI images from FBIRN (Function Biomedical Informatics Research Network) (Keator, 2016), COBRE (Center of Biomedical Research Excellence) (Çetin, 2014), BSNIP (Bipolar and Schizophrenia Network for Intermediate Phenotypes) (Tamminga, 2014), ABIDE (Autism Brain Imaging Data Exchange, release 1.0) (Di Martino, 2014), OASIS (Open Access Series of Imaging Studies, release 3.0) (Rubin, 1998), and HCP (Human Connectome Project, 1200 subjects release) (Van Essen, 2013).
Results:
In the 5-fold cross validated experiments BrainDynaMo showed competitive classification performance (Figure 2a), often outperforming other recent models for fMRI analysis with the exception of the larger HCP dataset. In our group analysis of the model-derived FNCs we clustered the matrices and compared the relative frequencies of found centroid FNCs across groups. The results of such analysis on the example of FBIRN data is shown in FIgure 2b. We believe that further analysis of the outlier centroid FNCs can reveal more information on the group differences and the mechanisms behind them.
Conclusions:
The presented BrainDynaMo model can dynamically learn functional network connectivity representations from fMRI data, offering a powerful tool for analyzing brain dynamics. The learned FNCs can be utilized in group analysis and, in the end, the knowledge discovery on the brain disorders and brain mechanics in general.
Modeling and Analysis Methods:
Classification and Predictive Modeling
fMRI Connectivity and Network Modeling 1
Methods Development 2
Task-Independent and Resting-State Analysis
Keywords:
Machine Learning
Modeling
MRI
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Are you Internal Review Board (IRB) certified? Please note: Failure to have IRB, if applicable will lead to automatic rejection of abstract.
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using APA citation style.
2. Di Martino, A. (2014). The autism brain imaging data exchange: towards a large-scale evaluation of the intrinsic brain architecture in autism. Molecular Psychiatry, 19 (6), 659–667.
3. Keator, D.B. (2016). The function biomedical informatics research network data repository. NeuroImage, 124, 1074–1079.
4. Lindquist, M. A. (2008). The Statistical Analysis of fMRI Data. Statistical Science, 23 (4), 439–464.
5. Rubin, E.H. (1998). A prospective study of cognitive function and onset of dementia
in cognitively healthy elders. Archives of neurology, 55 (3), 395–401.
6. Sharma, C. M. (2024). Diagnosis of mental disorders using machine learning: Literature review and bibliometric mapping from 2012 to 2023. Heliyon, 10(12), e32548.
7. Tamminga, C.A. (2014). Bipolar and schizophrenia network for intermediate phenotypes: Outcomes across the psychosis continuum. Schizophrenia Bulletin, 40 (Suppl_2), S131–S137.
8. Van Essen, D.C. (2013). The WU-Minn human connectome project: An overview. NeuroImage, 80, 62–79.
9. Yeung, H.W. (2023). Predicting sex, age, general cognition and mental health with machine learning on brain structural connectomes. Human Brain Mapping, 44 (5), 1913–1933.