Learning a Compact, Parcel-independent Representation of the fMRI Functional Connectivity
Keywords: dimensionality reduction, fMRI, variational autoencoder, performance evaluation, application
TL;DR: We show that learning a low-dimensional representation of fMRI functional connectivity data using variational autoencoders improves computational efficiency and generalizes well across datasets, aiding reproducibility and open science goals.
Abstract: Functional connectivity in functional magnetic resonance imaging (fMRI) data is often calculated at the level of area parcels. Given the data's low-dimensional nature, we posit a substantial degree of redundancy in these representations. Moreover, establishing correspondence across different individuals poses a significant challenge in that framework. We hypothesize that learning a compact representation of the functional connectivity data without losing the essential structure of the original data is possible. Our analysis, based on various performance benchmarks, indicates that the pre-computed mapping to low-dimensional latent space learned from the functional connectivity of one dataset generalizes well to another with both linear and non-linear autoencoder-based methods. Notably, the latent space learned using a variational autoencoder represents the data more effectively than linear methods at lower dimensions (2 dimensions). However, at higher dimensions (32 dimensions), the differences between linear and nonlinear dimensionality reduction methods diminish, rendering the performance comparable to the parcel space representation with 333 dimensions. Our findings highlight the potential of employing an established transformation to obtain a low-dimensional latent representation in future functional connectivity research, thereby solving the correspondence problem across parcel definitions, promoting reproducibility, and supporting open science objectives.
Primary Area: applications to neuroscience & cognitive science
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 7308
Loading