SpaRG: Sparsely Reconstructed Graphs for Generalizable fMRI Analysis

TL;DR

SpaRG addresses the challenge of interpreting deep learning on rs-fMRI while achieving cross-site generalization by learning a sparse, generalizable set of functional connections. It combines a trainable sparse mask $\boldsymbol{\mathcal{M}}$, a self-supervised variational autoencoder $\mathcal{A}$, and a graph convolutional classifier $\mathcal{F}$ in an end-to-end framework, optimizing sparsification, reconstruction, and classification losses with labeled and unlabeled data. On ABIDE sex classification, SpaRG can occlude up to $99\%$ of the original connections without sacrificing, and often improving, accuracy across in-distribution and out-of-distribution sites, with robustness observed across 64×64 and 1024×1024 DiFuMo parcellations. Qualitative analyses reveal that the preserved connections align with known networks (visual, attention, default mode), supporting interpretability and suggesting that self-supervised sparsification yields stable biomarkers under acquisition shifts.

Abstract

Deep learning can help uncover patterns in resting-state functional Magnetic Resonance Imaging (rs-fMRI) associated with psychiatric disorders and personal traits. Yet the problem of interpreting deep learning findings is rarely more evident than in fMRI analyses, as the data is sensitive to scanning effects and inherently difficult to visualize. We propose a simple approach to mitigate these challenges grounded on sparsification and self-supervision. Instead of extracting post-hoc feature attributions to uncover functional connections that are important to the target task, we identify a small subset of highly informative connections during training and occlude the rest. To this end, we jointly train a (1) sparse input mask, (2) variational autoencoder (VAE), and (3) downstream classifier in an end-to-end fashion. While we need a portion of labeled samples to train the classifier, we optimize the sparse mask and VAE with unlabeled data from additional acquisition sites, retaining only the input features that generalize well. We evaluate our method - Sparsely Reconstructed Graphs (SpaRG) - on the public ABIDE dataset for the task of sex classification, training with labeled cases from 18 sites and adapting the model to two additional out-of-distribution sites with a portion of unlabeled samples. For a relatively coarse parcellation (64 regions), SpaRG utilizes only 1% of the original connections while improving the classification accuracy across domains. Our code can be found at github.com/yanismiraoui/SpaRG.

Figures (3)

• Figure 1: SpaRG: a sparse mask $\mathcal{M}$, variational autoencoder (VAE) $\mathcal{A}$ that reconstructs the sparse inputs, and graph convolutional network (GCN) classifier $\mathcal{F}$ are trained in an end-to-end fashion to learn a subset of robust and informative functional connections. We interleave supervised training of the GCN with self-supervised steps, where we optimize the sparsification and autoencoding losses.
• Figure 2: Balanced accuracy on the ID and OOD sites for different levels of occlusion.
• Figure 3: Functional connections preserved by SpaRG for two parcel granularities, mapped to the 17-network atlas yeo2011organization. Similar connections are preserved by both models, highlighting connectivity involving the visual and default mode networks.