Spherical Convolutions in microstructural data to regress cognition. From synthetic to real data.
Poster No:
1564
Submission Type:
Abstract Submission
Authors:
Gabriela Gomez Jimenez1, Reuben Dorent2, Demian Wassermann3
Institutions:
1INRIA, Paris-Saclay, Ile de france, 2INRIA, Paris, Ile de france, 3MIND Team, Inria Saclay, Université Paris-Saclay, Palaiseau, France, Palaiseau, France
First Author:
Co-Author(s):
Demian Wassermann
MIND Team, Inria Saclay, Université Paris-Saclay, Palaiseau, France
Palaiseau, France
MIND Team, Inria Saclay, Université Paris-Saclay, Palaiseau, France
Palaiseau, France
Introduction:
Understanding the link between brain structure and behavior (BSB) is challenging (Genon, 2022). Diffusion Magnetic Resonance Imaging (DMRI) helps explore this link by measuring water molecules' motion within brain tissue (Le Bihan, 2015). Processing the signals associated to each gradient amplitude and direction of the magnetic field (q-space) (Assemlal, 2009) gives information about the water flow at the voxel level. Recent studies highlight the potential of Deep Learning (DL) methods to decode cognitive information from microstructural data (Jiménez, 2024). Relying on previous works that showed BSB correlations (Jiménez, 2024; Menon, 2020), we aim to build a model for brain microstructural data that (1) accounts for data directionality in each voxel and (2) predicts cognition. This work introduces a novel DL framework combining a Spherical Convolutional Neural Network (SCNN) to capture microstructural features' direction and a Multi-Layer Perceptron (MLP) to predict cognitive performance from DMRI. Our framework is assessed on both synthetic and real DMRI datasets.
Methods:
We simulate synthetic DMRI data for 800 subjects, 3 gradient amplitudes (1000, 2000, 3000s/mm2) and 100 voxels each, using dmipy (Rutger, 2019). The signal was modeled as a linear combination of 3 microstructural compartments: free water (fw), represented by a Gaussian Phase sphere model; soma, using a ball model with a 15.3 μm diameter (Evard, 2012) and extra-cellular (EC) tissue, modeled with a Zeppelin model. The signal is computed as S = ΣfiSi with i = (fw, soma and EC), where the fractional contributions fi sum to one. Synthetic cognitive scores were generated as a quadratic-interaction function of the fractional contributions to simulate realistic scenarios since they are the only voxel-varying parameters influencing the signal. For real data, we analyzed the same 3 gradient amplitudes for 881 subjects from the Human Connectome Project (HCP) dataset, focusing on 536 voxels in the left insula, a region linked to working memory. Cognitive performance was measured using the Working Memory Accuracy (Menon, 2020). Our DL framework consisted of two models: 1) an SCNN with 3 layers to reduce the input signal by 8 with a von Mises-Fisher kernel to capture the directionality features of the signal and 2) an MLP with two linear layers (hidden dim.: 64) to synthesize the features and predict cognitive scores. Model performance was assessed using the explained variance (EV) score.
Results:
The proposed model shows excellent performance on the synthetic data, including on test data (0.995 EV score). In contrast, its performance on real data was limited, showing underfitting during testing (negative r2 score) despite learning during training (Figure 1). To evaluate the benefits of SCNN for directional data, we conducted an ablation study where an MLP was trained without the SCNN layers. Figure 2 shows that our SCNN-based approach outperforms the MLP, achieving higher EV scores on training (EV: 0.962 vs -1.278) and testing (EV: -0.253 vs -0.793) sets. The results highlight the advantage of adding topological structure into data analysis. However, the lower EV and negative R2 scores on real data underscore the complexity of brain microstructure in real-world datasets compared to synthetic data.
Conclusions:
The study goal was to design a model for brain microstructural data that accounts for data directionality in each voxel and predicts cognition. Our framework, which exploits SCNN outperformed the MLP baseline. In terms of cognitive prediction, the model successfully captured BSB relationships in synthetic data. However, the performance dropped on real DMRI data, highlighting the challenges posed by the variability and complexity of BSB relationships. These results demonstrate the need for incorporating the topological structure of DMRI data in DL models to improve performance in the high-dimensional tasks of BSB analysis.
Modeling and Analysis Methods:
Classification and Predictive Modeling
Diffusion MRI Modeling and Analysis 2
Methods Development 1
Keywords:
Cognition
Computational Neuroscience
Machine Learning
MRI
Other - diffusion 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?
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:
Provide references using APA citation style.
[2] Evrard, H. C., Forro, T., & Logothetis, N. K. (2012). Von Economo neurons in the anterior insula of the macaque monkey. Neuron, 74(3), 482-489.
[2] Genon, S., Eickhoff, S. B., & Kharabian, S. (2022). Linking interindividual variability in brain structure to behaviour. Nature Reviews Neuroscience, 23(5), 307–318. https://doi.org/10.1038/s41583-022-00576-6
[3] Jiménez, G. G., & Wassermann, D. (2024). Deep multivariate autoencoder for capturing complexity in brain structure and behaviour relationships. arXiv preprint, arXiv:2409.01638. https://doi.org/10.48550/arXiv.2409.01638
[4] Le Bihan, D., & Iima, M. (2015). Diffusion magnetic resonance imaging: What water tells us about biological tissues. PLoS Biology, 13(7), e1002203. https://doi.org/10.1371/journal.pbio.1002203
[5] Menon, V., Gallardo, G., Pinsk, M. A., Nguyen, V. D., Li, J. R., Cai, W., & Wassermann, D. (2020). Microstructural organization of human insula is linked to its macrofunctional circuitry and predicts cognitive control. eLife, 9, e53470.
[6] Rutger Fick, Demian Wassermann and Rachid Deriche, "The Dmipy Toolbox: Diffusion MRI Multi-Compartment Modeling and Microstructure Recovery Made Easy", Frontiers in Neuroinformatics 13 (2019): 64.