Reduce, Augment, Contrast : A Novel Approach For Mapping Brain-Cognition Associations
Poster No:
1536
Submission Type:
Abstract Submission
Authors:
Malo Renaudin1,2, Victoria Shevchenko1,2,3,4, Alexandre Le Bris1,2, R. Austin Benn3,4, Daniel Margulies3,4, Demian Wassermann1,2
Institutions:
1MIND Team, Inria Saclay, Université Paris-Saclay, Palaiseau, France, 2Neurospin, CEA, Gif-sur-Yvette, France, 3Cognitive Neuroanatomy Lab, Université Paris Cité, INCC UMR 8002, CNRS, Paris, France, 4Wellcome Centre for Integrative Neuroimaging, FMRIB Centre, Nuffield Department of Clinical Neurosciences, University of Oxford, Oxford, United Kingdom
First Author:
Malo Renaudin
MIND Team, Inria Saclay, Université Paris-Saclay|Neurospin, CEA
Palaiseau, France|Gif-sur-Yvette, France
MIND Team, Inria Saclay, Université Paris-Saclay|Neurospin, CEA
Palaiseau, France|Gif-sur-Yvette, France
Co-Author(s):
Victoria Shevchenko
MIND Team, Inria Saclay, Université Paris-Saclay|Neurospin, CEA|Cognitive Neuroanatomy Lab, Université Paris Cité, INCC UMR 8002, CNRS|Wellcome Centre for Integrative Neuroimaging, FMRIB Centre, Nuffield Department of Clinical Neurosciences, University of Oxford
Palaiseau, France|Gif-sur-Yvette, France|Paris, France|Oxford, United Kingdom
MIND Team, Inria Saclay, Université Paris-Saclay|Neurospin, CEA|Cognitive Neuroanatomy Lab, Université Paris Cité, INCC UMR 8002, CNRS|Wellcome Centre for Integrative Neuroimaging, FMRIB Centre, Nuffield Department of Clinical Neurosciences, University of Oxford
Palaiseau, France|Gif-sur-Yvette, France|Paris, France|Oxford, United Kingdom
Alexandre Le Bris
MIND Team, Inria Saclay, Université Paris-Saclay|Neurospin, CEA
Palaiseau, France|Gif-sur-Yvette, France
MIND Team, Inria Saclay, Université Paris-Saclay|Neurospin, CEA
Palaiseau, France|Gif-sur-Yvette, France
R. Austin Benn
Cognitive Neuroanatomy Lab, Université Paris Cité, INCC UMR 8002, CNRS|Wellcome Centre for Integrative Neuroimaging, FMRIB Centre, Nuffield Department of Clinical Neurosciences, University of Oxford
Paris, France|Oxford, United Kingdom
Cognitive Neuroanatomy Lab, Université Paris Cité, INCC UMR 8002, CNRS|Wellcome Centre for Integrative Neuroimaging, FMRIB Centre, Nuffield Department of Clinical Neurosciences, University of Oxford
Paris, France|Oxford, United Kingdom
Daniel Margulies
Cognitive Neuroanatomy Lab, Université Paris Cité, INCC UMR 8002, CNRS|Wellcome Centre for Integrative Neuroimaging, FMRIB Centre, Nuffield Department of Clinical Neurosciences, University of Oxford
Paris, France|Oxford, United Kingdom
Cognitive Neuroanatomy Lab, Université Paris Cité, INCC UMR 8002, CNRS|Wellcome Centre for Integrative Neuroimaging, FMRIB Centre, Nuffield Department of Clinical Neurosciences, University of Oxford
Paris, France|Oxford, United Kingdom
Demian Wassermann
MIND Team, Inria Saclay, Université Paris-Saclay|Neurospin, CEA
Palaiseau, France|Gif-sur-Yvette, France
MIND Team, Inria Saclay, Université Paris-Saclay|Neurospin, CEA
Palaiseau, France|Gif-sur-Yvette, France
Introduction:
Predicting cognition from functional connectomes (FC) is hindered by high-dimensional neuroimaging data and limited sample sizes (Helmer et al. 2024). In this study we combined unsupervised dimensionality reduction, data augmentation and supervised contrastive learning to overcome these obstacles and predict cognitive scores (CS) from FC, achieving a median 9% error margin across CS on average across trials.
We reduce the dimensionality of FC matrices by taking advantage of their mathematical structure. Critically, FC lies on the Riemannian manifold of Symmetric Positive Definite (SPD) matrices, we leveraged this property through an SPD autoencoder (D'Souza et al. 2021).To tackle sample size limitation, we introduced three augmentation strategies tailored for the obtained reduced FC.
Finally, to perform supervised multivariate prediction of CS, we employed contrastive learning (CL). CL aligns reduced FC representations that are similar in terms of CS in the embedding space (Wang and Isola, 2022), which improves the embeddings' ability to capture CS variance. Notably, while the majority of contrastive learning losses are designed for classification tasks, we utilized a variant adapted to regression (Barbano et al. 2023).
We reduce the dimensionality of FC matrices by taking advantage of their mathematical structure. Critically, FC lies on the Riemannian manifold of Symmetric Positive Definite (SPD) matrices, we leveraged this property through an SPD autoencoder (D'Souza et al. 2021).To tackle sample size limitation, we introduced three augmentation strategies tailored for the obtained reduced FC.
Finally, to perform supervised multivariate prediction of CS, we employed contrastive learning (CL). CL aligns reduced FC representations that are similar in terms of CS in the embedding space (Wang and Isola, 2022), which improves the embeddings' ability to capture CS variance. Notably, while the majority of contrastive learning losses are designed for classification tasks, we utilized a variant adapted to regression (Barbano et al. 2023).
Methods:
We analyzed resting-state fMRI data from the CamCAN dataset (Taylor et al. 2017) for 352 subjects. We preprocessed fMRI data with fMRIPrep (Esteban et al. 2019). For each parcel in the Schaefer parcellation (400 parcels), we computed the mean BOLD time-series, followed by pairwise Pearson correlation to get connectivity matrices. We preprocessed CS as outlined in Engemann et al. (2020).
To predict CS, we employed a multi-stage model trained on 246 subjects. First, we pretrained an SPD autoencoder (Fig.1A) with a norm loss (Fig.1A, eq.3) to reduce (Fig. 1A, eq. 1) and reconstruct (Fig. 1A, eq. 2) FC matrices. We applied three augmentation strategies on the reduced FC (Fig.1B). Those included adding structured noise using singular value decomposition (SVD) (Fig.1B, eq.1), value randomization (Fig.1B, eq.2), and Graph Path Convolution (GPC) (Li et al. 2022) (Fig.1B, eq.3). Next, we vectorized reduced FC matrices and processed them with a multilayer perceptron (MLP), trained with a supervised contrastive loss for regression (Fig.1C, eq.1 and 2) to obtain cognitively-informed embeddings. We supplemented this loss with a direction regularization term (Mohan et al. 2020) (Fig.1C, eq.3). Notably, we trained our model in a multivariate setting, projecting embeddings with respect to both CS of interest and confounding variables (e.g. age and cardiovascular measures). We obtained final predictions through an additional MLP, trained with Mean Squared Error (MSE) (Fig.1D, eq.1) loss to decode CS from the behavioral embeddings.
To predict CS, we employed a multi-stage model trained on 246 subjects. First, we pretrained an SPD autoencoder (Fig.1A) with a norm loss (Fig.1A, eq.3) to reduce (Fig. 1A, eq. 1) and reconstruct (Fig. 1A, eq. 2) FC matrices. We applied three augmentation strategies on the reduced FC (Fig.1B). Those included adding structured noise using singular value decomposition (SVD) (Fig.1B, eq.1), value randomization (Fig.1B, eq.2), and Graph Path Convolution (GPC) (Li et al. 2022) (Fig.1B, eq.3). Next, we vectorized reduced FC matrices and processed them with a multilayer perceptron (MLP), trained with a supervised contrastive loss for regression (Fig.1C, eq.1 and 2) to obtain cognitively-informed embeddings. We supplemented this loss with a direction regularization term (Mohan et al. 2020) (Fig.1C, eq.3). Notably, we trained our model in a multivariate setting, projecting embeddings with respect to both CS of interest and confounding variables (e.g. age and cardiovascular measures). We obtained final predictions through an additional MLP, trained with Mean Squared Error (MSE) (Fig.1D, eq.1) loss to decode CS from the behavioral embeddings.
Results:
We quantify FC reconstruction quality with mean absolute percentage error (MAPE) (Fig. 2A, panel 1). These matrices show a structured error, indicating that our autoencoder reconstruction maintains the organization of connectivity across subjects. Additionally, performance is stable across regions (Fig. 2A, panel 3).
We compare our method with CCA, the standard multivariate analysis for this task (Helmer et al. 2024). Our model achieves better correlation with less dispersion (30±3% vs 28±15%) (Fig 2B). With respect to the MAPE of CS values, we have 9±0.3% versus 125±34% for CCA, which implies that we can predict CS within a median 9% error margin across CS on average across trials.
We compare our method with CCA, the standard multivariate analysis for this task (Helmer et al. 2024). Our model achieves better correlation with less dispersion (30±3% vs 28±15%) (Fig 2B). With respect to the MAPE of CS values, we have 9±0.3% versus 125±34% for CCA, which implies that we can predict CS within a median 9% error margin across CS on average across trials.
Conclusions:
Our method preserves the SPD structure of FC during reconstruction. It also retains sufficient variance for strong CS decoding performance from the obtained embeddings. This approach holds promise for unsupervised dimensionality reduction combined with supervised multivariate cognition prediction from FC.
Higher Cognitive Functions:
Higher Cognitive Functions Other
Modeling and Analysis Methods:
Classification and Predictive Modeling
Connectivity (eg. functional, effective, structural)
Methods Development 1
Multivariate Approaches 2
Keywords:
Computational Neuroscience
FUNCTIONAL MRI
Machine Learning
Multivariate
Other - CCA
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?
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.
D’Souza, N. S. et al. (2021). A Matrix Autoencoder Framework to Align the Functional and Structural Connectivity Manifolds as Guided by Behavioral Phenotypes (arXiv:2105.14409). arXiv. https://doi.org/10.48550/arXiv.2105.14409
Engemann, D. A. et al. (2020). Combining magnetoencephalography with magnetic resonance imaging enhances learning of surrogate-biomarkers. eLife, 9, e54055. https://doi.org/10.7554/eLife.54055
Esteban, O. et al. (2019). fMRIPrep: A robust preprocessing pipeline for functional MRI. Nature Methods, 16(1), 111–116. https://doi.org/10.1038/s41592-018-0235-4
Helmer, M. et al. (2024). On the stability of canonical correlation analysis and partial least squares with application to brain-behavior associations. Communications Biology, 7(1), 1–15. https://doi.org/10.1038/s42003-024-05869-4
Li, Y. et al. (2022). Brain Connectivity Based Graph Convolutional Networks and Its Application to Infant Age Prediction. IEEE Transactions on Medical Imaging, 41(10), 2764–2776. https://doi.org/10.1109/TMI.2022.3171778
Mohan, D. et al. (2020). Moving in the Right Direction: A Regularization for Deep Metric Learning. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 14579–14587. https://doi.org/10.1109/CVPR42600.2020.01460
Taylor, J. R. et al. (2017). The Cambridge Centre for Ageing and Neuroscience (Cam-CAN) data repository: Structural and functional MRI, MEG, and cognitive data from a cross-sectional adult lifespan sample. NeuroImage, 144, 262–269. https://doi.org/10.1016/j.neuroimage.2015
Wang, T., & Isola, P. (2022). Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hypersphere (arXiv:2005.10242). arXiv. https://doi.org/10.48550/arXiv.2005.10242