Beyond Site Agreement: Re-estimation for Brain Network Generalization

Yingxu Wang; Kunyu Zhang; Yanwu Yang3; Thomas Wolfers; Yujie Wu; Siyang Gao; Nan Yin

cs.LGSep 28, 2026

Beyond Site Agreement: Re-estimation for Brain Network Generalization

Authors: Yingxu Wang, Kunyu Zhang, Yanwu Yang3, Thomas Wolfers, Yujie Wu, Siyang Gao, Nan Yin

Organizations: Mohamed bin Zayed University of Artificial Intelligence · Zhengzhou University · University Hospital Tübingen · The Hong Kong Polytechnic University · City University of Hong Kong · The Education University of Hong Kong

Abstract

Cross-site out-of-distribution (OOD) generalization in resting-state functional magnetic resonance imaging (rs-fMRI) often relies on learning task-discriminative representations from full-scan functional connectivity (FC) graphs and promoting invariance across source sites. However, FC graphs are estimated from finite, temporally correlated blood-oxygen-level-dependent (BOLD) sequences. Cross-site agreement therefore does not necessarily imply that predictive evidence remains supported under FC re-estimation within the same scan. In this paper, we propose Brain Network Re-estimation-Informed OOD Learning (BRIO), a framework that uses within-scan FC re-estimation to guide cross-site alignment. BRIO maps fullscan graphs and their re-estimates into consistently indexed connectome factors, enabling comparisons of their predictive contributions. It assesses re-estimation support from changes in these contributions relative to within-class subject variability and class separation. For each source-site pair and class, this task-calibrated support from both sites is combined with predictive relevance to form pairwise qualifications, which determine relative factor weights and overall alignment strength. Leave-one-site-out experiments on four real-world datasets (ABIDE, REST-metaMDD, SRPBS, and ABCD) show that BRIO consistently outperforms competitive baselines, with relative improvements of up to 3.8% in accuracy. These gains also persist under an alternative brain parcellation on ABIDE.

Figures & tables

Figure 1: Overview of BRIO. Block resampling of the same BOLD scan produces FC re-estimates. Shared factorization uses a shared GNN and ROI map to obtain consistently indexed factors from full-scan and re-estimated graphs. Predictive relevance and task-calibrated re-estimation support from both sites form pairwise qualifications for each site pair and class, which control shared weighting and alignment strength for cross-site alignment of reweighted full-scan features.

Type	Method	ABIDE		ABIDE (CC200)		REST-meta-MDD		SRPBS		ABCD (ADHD-task)
Type	Method	AUC	ACC	AUC	ACC	AUC	ACC	AUC	ACC	AUC	ACC
GNN	GCN	61.91 ±10.32	56.12 ±9.66	50.21 ±8.86	48.38 ±8.55	60.58 ±4.18	56.26 ±2.87	65.24 ±19.47	61.68 ±19.17	71.75 ±6.52	63.14 ±9.15
	GAT	58.21 ±8.69	52.31 ±6.01	57.68 ±9.37	54.69 ±7.74	59.97 ±9.38	56.17 ±6.72	66.68 ±15.58	60.86 ±14.38	70.37 ±11.32	63.32 ±11.20
	GIN	57.73 ±5.21	52.08 ±3.10	52.40 ±8.83	47.84 ±10.23	56.55 ±6.72	53.46 ±4.72	63.74 ±21.68	60.19 ±18.08	68.97 ±8.84	60.73 ±7.03
OOD	CORAL	55.48 ±6.12	50.23 ±5.49	53.75 ±10.69	50.06 ±7.42	60.37 ±8.97	57.18 ±7.78	63.88 ±15.48	59.86 ±12.68	69.43 ±7.70	62.55 ±10.46
OOD	IRM	58.56 ±6.12	52.98 ±3.85	51.77 ±10.64	50.84 ±6.99	58.56 ±4.38	54.18 ±3.35	60.98 ±21.87	59.27 ±16.78	70.37 ±6.60	59.06 ±7.57

Table 1: Results on ABIDE, ABIDE (CC200), REST-meta-MDD, SRPBS, and ABCD (ADHD-task) over LOSO sites. Bold results indicate the best performance.

Panel of Figure 2 — Figure 2: Ablation results in (a,b) and sensitivity results in (c,d) on ABIDE and REST-meta-MDD.

Panel of Figure 3 — Figure 3: Cross-method relationships (a), alignment trajectories (b), mismatch rates (c), and factor retention (d) on ABIDE.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Description
$\mathcal{E}_{\mathrm{tr}},\mathcal{E}_{\mathrm{te}},E$	Source and unseen test site sets, and the number of source sites.
$\bm{X}_{i}^{e},Y_{i}^{e}$	Preprocessed BOLD sequence and binary label of subject $i$ at site $e$ .
$T_{i}^{e},P$	Numbers of valid time points and ROIs.
$\mathcal{I}_{e,c},N_{e,c}$	Subject index set for site $e$ and class $c$ , and its size.
$\mathcal{B}_{e,c}$	Training mini-batch subset for site $e$ and class $c$ .
$\Psi,\mathcal{R}_{k},K$	FC estimator, joint circular moving-block resampling operator, and number of FC re-estimates per subject.

Appendix

Table 2: Summary of key notations.

Dataset	Task	Samples	Class counts	Parcellation	Time points
ABIDE	ASD vs. TD	1,025	488 / 537	AAL116; CC200	$194.51\pm 58.58$
REST-meta-MDD	MDD vs. HC	2,428	1,300 / 1,128	AAL116	$207.54\pm 29.29$
SRPBS	MDD vs. HC	1,000	500 / 500	AAL116	107–284
ABCD	ADHD vs. HC	425	213 / 212	AAL116	383

Appendix

Table 3: Statistics of the datasets.

Type	Method	ABIDE		ABIDE (CC200)		REST-meta-MDD		SRPBS		ABCD (ADHD-task)
Type	Method	AUC	ACC	AUC	ACC	AUC	ACC	AUC	ACC	AUC	ACC
GNN	GCN	63.45 ±4.12	59.21 ±3.85	51.88 ±4.02	51.05 ±3.74	62.15 ±2.88	57.02 ±2.45	84.12 ±3.85	75.34 ±4.12	72.18 ±5.12	64.05 ±4.88
	GAT	59.12 ±3.95	55.44 ±3.55	58.62 ±4.15	57.18 ±3.82	63.84 ±3.12	59.12 ±2.65	81.75 ±4.05	70.88 ±4.35	71.02 ±5.45	63.88 ±5.10
	GIN	58.74 ±3.65	54.95 ±3.22	53.15 ±4.33	50.42 ±3.95	59.33 ±2.54	56.14 ±2.22	82.05 ±3.78	71.45 ±3.92	69.45 ±5.33	61.22 ±4.75
OOD	CORAL	56.88 ±4.05	53.25 ±3.75	55.14 ±4.45	53.75 ±4.05	59.95 ±3.15	57.85 ±2.85	81.14 ±4.12	70.82 ±4.05	70.12 ±5.25	63.45 ±4.85
OOD	IRM	60.15 ±3.88	56.78 ±3.45	53.42 ±4.65	54.12 ±4.11	60.75 ±2.95	57.44 ±2.75	80.22 ±3.94	70.15 ±3.88	71.25 ±5.08	60.14 ±4.95

Appendix

Table 4: ID results on ABIDE, ABIDE (CC200), REST-meta-MDD, SRPBS, and ABCD (ADHD-task) under ten-fold cross-validation stratified by site and class, reported as mean

\pm

standard deviation across test folds. Bold results indicate the best performance.

Panel of Figure 4 — Figure 4: Cross-method relationships (a), alignment trajectories (b), mismatch rates (c), and factor retention (d) on REST-meta-MDD.

Panel of Figure 5 — Figure 5: Cross-method relationships (a), alignment trajectories (b), mismatch rates (c), and factor retention (d) on SRPBS.

Panel of Figure 6 — Figure 6: Cross-method relationships (a), alignment trajectories (b), mismatch rates (c), and factor retention (d) on ABCD.

Panel of Figure 7 — Figure 7: Ablation study on ABIDE (CC200), SRPBS, and ABCD.

Panel of Figure 8 — Figure 8: Sensitivity study on ABIDE (CC200), SRPBS, and ABCD.

Dataset	$0.5\times$			$1\times$ (Default)			$2\times$
Dataset	$\ell_{b}$	AUC	ACC	$\ell_{b}$	AUC	ACC	$\ell_{b}$	AUC	ACC
ABIDE	3 (6.0)	65.84 ±4.67	60.25 ±4.15	6 (12.0)	67.53 ±4.12	62.18 ±3.68	12 (24.0)	66.71 ±4.28	61.54 ±3.92
ABIDE (CC200)	3 (6.0)	69.45 ±5.25	60.12 ±4.58	6 (12.0)	70.86 ±4.87	61.92 ±3.94	12 (24.0)	68.72 ±5.15	59.88 ±4.25
REST-meta-MDD	4 (8.0)	67.85 ±4.22	63.38 ±3.42	7 (14.0)	68.84 ±3.75	64.42 ±2.91	14 (28.0)	67.62 ±3.82	63.85 ±3.18
SRPBS	3 (6.0)	81.35 ±6.35	76.54 ±5.82	6 (12.0)	83.91 ±5.63	78.82 ±5.28	12 (24.0)	83.72 ±5.58	78.45 ±5.15
ABCD (ADHD-task)	7 (5.6)	72.15 ±5.45	65.12 ±4.85	15 (12.0)	75.98 ±4.39	68.74 ±3.82	30 (24.0)	75.45 ±4.52	67.92 ±3.95

Appendix

Table 5: Sensitivity of BRIO to moving-block length

\ell_{b}

. Bold indicates the highest mean.

Panel of Figure 9 — Figure 9: Performance under reduced scan data on different datasets.

Method	ABIDE		REST-meta-MDD		SRPBS		ABCD
Method	Time	Memory	Time	Memory	Time	Memory	Time	Memory
GSAT	0.68	2.1	1.52	3.3	0.48	2.0	0.34	1.7
DisC	1.05	2.4	2.38	3.8	0.74	2.2	0.51	1.9
CEPG	0.58	2.0	1.34	3.1	0.41	1.9	0.28	1.6
DiSCO	1.65	2.9	3.85	4.6	1.18	2.7	0.78	2.3
BrainNetTF	0.78	2.0	1.65	3.2	0.49	1.9	0.32	1.6

Appendix

Table 6: Training time (in seconds) and GPU memory consumption (in GB) for each training epoch.

Panel of Figure 10 — Figure 10: Anatomical distribution of qualification-weighted ROI profiles on ABIDE and REST-meta-MDD.

Explore similar work

May 7, 2026cs.LG

Learning beyond Site Bias for OOD Generalization in Brain Networks

Graph-based learning from functional magnetic resonance imaging (fMRI) has shown strong potential for brain network analysis. However, existing methods often degrade under cross-site out-of-distribution (OOD) settings, as site-dependent confounder effects can obscure disease-related connectivity patterns and static functional connectivity (FC) does not explicitly capture informative within-scan variations. In this paper, we propose Cross-site OOD Robust brain nEtwork (CORE), a unified framework for brain network learning across unseen sites. First, CORE estimates site-specific confounder effects and aggregates the resulting deconfounders by cross-source reliability for unseen-site correction, while extracting a population scaffold of reproducible label-associated connections from site-wise residualized FC. It then summarizes temporal variations on scaffold edges into compact descriptors for line-graph modeling. Finally, prior-guided subject-adaptive gating modulates message passing, balancing population priors with individual variability. Extensive leave-one-site-out experiments on ABIDE, REST-meta-MDD, SRPBS, and ABCD demonstrate that CORE consistently outperforms competitive baselines, with up to a 10.3% relative improvement in accuracy. These gains also persist across different brain parcellation schemes on ABIDE.

Yingxu Wang, Kunyu Zhang, Yanwu Yang +5

Zhengzhou University · City University of Hong Kong

Human Brain Networks Functional Connectivity Graph Representation Learning Beyond Retrieval Prior Knowledge Integration

Ask PDF

Aug 31, 2026eess.SP

Benchmarking External Generalization of SPD Matrix Learning for Resting-State fMRI Connectome Prediction

Resting-state functional magnetic resonance imaging (rs-fMRI) functional connectivity (FC) matrices are widely used for individual-level prediction, but strong performance within one cohort may not generalize to a new cohort. We ask whether within-dataset performance remains when the test data come from an entirely held-out rs-fMRI dataset. Each scan is represented as a regularized symmetric positive definite (SPD) correlation connectome, which allows methods to use the geometry of the SPD manifold. We introduce a reproducible age-prediction benchmark across six rs-fMRI datasets: COBRE, ADNIDOD, Cam-CAN, ABIDE, OASIS-3, and ADNI. The benchmark compares a vectorized correlation baseline, Tangent-Space Ridge, SPDNet, and split-wise Riemannian harmonization under within-dataset GroupKFold, pooled GroupKFold, and leave-one-dataset-out (LODO) evaluation. Within-dataset and pooled GroupKFold results are substantially more favorable than LODO results. When an entire dataset is held out, prediction error increases, differences among methods narrow, and performance is strongly affected by age-range mismatch and cohort heterogeneity. The benchmark provides common inputs, model settings, data splits, and analysis scripts so that future SPD matrix learning methods can be evaluated under the same external-validation protocol.

Ce Ju, Antoine Collas, Florent Bouchard +1

Inria, CEA, Universit´e Paris-Saclay, Palaiseau, France · Universit´e Paris-Saclay, CNRS, CentraleSup´elec, L2S, France

Human Connectome Project Full-Rank Correlation Matrices Vascular Age Kernel Ridge Regression Matrix

Ask PDF

Jun 28, 2026cs.LG

BrainRiem: Riemannian Prototype Learning for Source-Free Cross-Site Brain Network Diagnosis

Multi-site functional MRI (fMRI) studies are essential for robust neuropsychiatric diagnosis yet suffer severe domain shifts from scanner heterogeneity, demographics, and site-specific acquisition protocols. Traditional domain adaptation requires concurrent source and target data access, violating clinical privacy regulations. Moreover, functional connectivity matrices lie on the Symmetric Positive Definite (SPD) manifold, where Euclidean operations cause geometric distortions corrupting diagnostic patterns. We propose BrainRiem, a source-free domain adaptation framework learning compact Riemannian brain prototypes via manifold-aware bi-level optimization. It employs the Log-Euclidean Metric to ensure prototypes remain valid SPD matrices, while Dirichlet Energy spectral calibration aligns their frequency characteristics with real brain networks. Only anonymized prototypes are transmitted to target sites, serving as stable anchors for training local models without source data access and reducing leakage under the evaluated attacks. Comprehensive experiments on ABIDE and REST-meta-MDD show BrainRiem consistently outperforms state-of-the-art source-free, traditional, and graph domain adaptation methods across diverse scanners and demographics. Notably, learned prototypes exhibit biologically interpretable connectivity patterns aligning with established neuroscience findings, validating the necessity of Riemannian geometry for brain network analysis.

Kunyu Zhang, Tianxiang Xu

Zhengzhou University, Zhengzhou 450000, China · Peking University, Beijing 102600, China

Human Brain Networks Domain Adaptation Riemannian Manifolds

Ask PDF