Title : Interpretable domain generalization for multi-site functional connectivity analysis
Abstract:
Multi-site functional MRI provides an important foundation for developing reproducible neuroimaging models across diverse populations and clinical settings. However, differences in scanners, acquisition protocols, preprocessing procedures, and participant characteristics can produce substantial variation in functional connectivity (FC) across imaging sites. As a result, models developed using data from existing institutions may perform poorly when applied to a new clinic, cohort, or scanner.
Current approaches do not fully address this challenge. Many domain-adaptation methods use data from the future test site during model development. However, a model may need to be finalized before data from a new institution are available. In addition, many existing methods produce dense, low-dimensional representations that are difficult to relate back to specific brain connections.
We propose a Scale-Aware Row-Sparse Domain Generalization framework (SA-RSA), a source-only domaingeneralization approach designed for predicting unseen data. SA-RSA accounts for differences in the reliability of individual training sites, reduces both overall and outcome-specific differences among sites, preserves biologically relevant group separation, and identifies a compact set of informative ROI-ROI connections. The resulting model is fully trained using existing source sites and applied to a new site without retraining or target-site adaptation.
We evaluate SA-RSA on three independent multi-site resting-state fMRI datasets: ABIDE I, ABIDE II, and ADHD-200. Autism spectrum disorder (ASD) and attention-deficit/hyperactivity disorder (ADHD) classification are used as challenging neuroimaging applications. Under a strict leave-one-site-out design, ABIDE I is used for primary method development, ABIDE II provides independent replication for ASD classification, and ADHD-200 is used to evaluate generalization to a different neurodevelopmental disorder. SA-RSA achieved mean classification accuracies of 71.60% on ABIDE I, 68.36% on ABIDE II, and 62.04% on ADHD-200, outperforming the strongest comparison methods by 5.03, 5.23, and 9.31 percentage points, respectively.
The results demonstrate that SA-RSA improves unseen-site generalization across independent multi-site neuroimaging datasets and across two neurodevelopmental disorders while retaining an interpretable functional connectivity representation. By jointly accounting for cross-site heterogeneity, preserving diagnostically discriminative information, and identifying transferable ROI-ROI connections, SA-RSA provides a promising framework for robust and interpretable multi-site neuroimaging analysis. The longer-term goal is to support interpretable neuroimaging models that can be developed at existing institutions and applied more reliably to sites that were unavailable during model development.
Keywords: Domain generalization; functional connectivity; resting-state fMRI; autism spectrum disorder; attention-deficit/hyperactivity disorder; multi-site neuroimaging; interpretable machine learning
