PaperlayerView exact revision on arXiv

Source-linked AI summary

Beyond Site Agreement: Re-estimation for Brain Network Generalization

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

arXiv:2609.34611v1cs.LG

TL;DR

Cross-site rs-fMRI models seek representations that transfer across sites, but cross-site agreement alone does not show whether predictive evidence remains supported under FC re-estimation within a scan. BRIO uses task-calibrated re-estimation support to guide cross-site alignment. In leave-one-site-out experiments, it consistently outperforms competitive baselines across four datasets, with relative accuracy gains of up to 3.8%.

  • Problem

    Cross-site OOD methods seek task-discriminative, site-invariant representations, but full-scan cross-site agreement does not reveal whether predictive evidence remains supported under within-scan FC re-estimation.

    Source: §1 · pp. 1–2

  • Method

    BRIO maps full-scan and re-estimated graphs to consistently indexed connectome factors, then combines predictive relevance with task-calibrated support from both sites to set factor weights and alignment strength.

    Source: §1 · p. 2, §3.3 · p. 7

  • Results

    BRIO consistently outperforms competitive baselines in leave-one-site-out experiments on ABIDE, REST-metaMDD, SRPBS, and ABCD, with relative accuracy gains of up to 3.8%.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The findings support estimation stability as a criterion for which factors to align, rather than considering only how much to align.

    Source: §5 · p. 10

    The setup assumes sites share the label space, preprocessing pipeline, parcellation, and FC estimator, although their joint distributions may differ.

    Source: §3 · p. 3

Abstract

from arXiv · show

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.

1 INTRODUCTION

Cross-site rs-fMRI models seek task-discriminative, site-invariant representations, but agreement across sites does not establish that predictive evidence persists when FC is re-estimated within a scan. BRIO uses task-calibrated re-estimation support to guide which factors are aligned and how strongly.

1 INTRODUCTION

  • GNNs use connections among brain regions to learn predictive subject representations from FC graphs built from ROI-level BOLD time series.The paper situates this approach in non-invasive rs-fMRI characterization of macroscale functional interactions.Source: §1 · p. 1
  • FC estimates from finite, temporally correlated BOLD sequences can change under within-scan re-estimation, so low cross-site discrepancy alone does not establish that predictive evidence remains supported.Re-estimation changes are within-subject changes and are not directly characterized by discrepancies between site-level representation distributions.Source: §1 · pp. 1–2
  • The paper frames three challenges: probing variability from one scan while retaining temporal dependence, comparing pattern-level contributions consistently, and interpreting changes in both sites’ task contexts.A graph-level readout aggregates multiple connectivity patterns, while the same contribution change can have different implications across sites with different within-class variability and class separation.Source: §1 · p. 2
  • BRIO consistently outperforms competitive baselines across four datasets, with relative accuracy gains of up to 3.8% and gains maintained under an alternative brain parcellation.The experiments use leave-one-site-out evaluation on ABIDE, REST-metaMDD, SRPBS, and ABCD.Source: §1 · p. 2
  • BRIO maps full-scan and re-estimated graphs into consistently indexed connectome factors, then combines predictive relevance with task-calibrated support from both sites to set factor weights and class-conditional alignment strength.Support is calibrated against within-class subject variability and class separation for each source-site pair and class.Source: §1 · p. 2

2 RELATED WORK

Related work addresses cross-site distribution shifts through graph OOD learning, neuroimaging harmonization, and transferable brain-network representations. Connectome reliability research studies finite-sample estimation, while BRIO uses within-scan re-estimation support to qualify factors for alignment.

2 RELATED WORK

  • Graph OOD approaches target invariant subgraphs, separation of invariant and spurious information, and representation alignment; neuroimaging approaches also use harmonization and transferable brain-network representations.Related strategies include feature selection, causal augmentation, and site-aware deconfounding with transferable connectivity dynamics.Source: §2 · pp. 2–3
  • BRIO’s overview links same-scan block-resampled FC estimates to shared factors, then uses site-pair and class qualifications to control factor weighting and cross-site alignment strength.The framework’s related-work distinction is its use of within-scan re-estimation support to qualify predictive factors for alignment.Source: Figure 1 · p. 3, §2 · p. 3
  • Connectome reliability work considers effects of scan duration, temporal sampling, and autocorrelation, alongside test–retest studies, estimator comparisons, bootstrap procedures, and uncertainty modeling.Other lines learn connectome representations from BOLD signals or use temporal segments for contrastive learning.Source: §2 · p. 3

3 METHODOLOGY

BRIO evaluates predictive contributions across full-scan FC graphs and within-scan re-estimates, then uses predictive relevance and re-estimation support to guide class-conditional cross-site alignment.

3.4 LEARNING OBJECTIVE

  • BRIO combines source-site classification with qualification-guided alignment of full-scan factor representations across source-site pairs and classes.The overall objective includes a classification loss balanced across source sites and classes, alongside the alignment loss.Source: §3 · p. 4, §3.3 · pp. 6–7, §3.4 · p. 7

3.1 CORRESPONDENCE-PRESERVING CONNECTOME FACTORIZATION

  • Jointly resampled temporal blocks generate within-scan FC re-estimates while retaining local dependence and cross-ROI temporal correspondence.A shared signed GNN and shared ROI assignment map full-scan and re-estimated graphs into consistently indexed factors for comparison.Source: §3 · p. 3, §3.1 · pp. 4–5

3.2 TASK-CALIBRATED RE-ESTIMATION QUALIFICATION

  • Task-calibrated qualification combines full-scan predictive relevance with re-estimation support measured relative to a local reference scale.The reference uses within-class variability and class separation; higher support indicates smaller mean squared contribution changes relative to that scale.Source: §3.2 · pp. 5–6
  • Proposition 1 shows that preserving a factor’s class-contrast sign requires less support as its full-scan class separation increases.The required support decreases from one for weakly separated factors to one half for strongly separated factors, motivating joint qualification by relevance and support.Source: §3.2 · p. 6

3.3 QUALIFICATION-GUIDED CROSS-SITE ALIGNMENT

  • For each source-site pair and class, BRIO weights factor qualifications using predictive relevance and support from both sites, with their sum setting alignment strength.The geometric mean treats sites symmetrically and decreases when either site’s support decreases.Source: §3.3 · pp. 6–7

4 EXPERIMENTS

BRIO is evaluated in leave-one-site-out experiments, component ablations, sensitivity analyses, and ABIDE case studies. The results test its performance and how re-estimation qualifications relate to cross-site alignment and factor transferability.

4 EXPERIMENTS

  • Leave-one-site-out experiments cover ABIDE, REST-meta-MDD, SRPBS, ABCD, and ABIDE (CC200), comparing BRIO with general GNN, OOD, graph OOD, and brain-network methods.The datasets cover ASD, MDD, and ADHD prediction tasks; ABCD is a developmental cohort.Source: §4 · p. 8, §4.1 · p. 8

4.1 PERFORMANCE COMPARISON

  • Under LOSO on ABIDE, REST-meta-MDD, SRPBS, and ABCD, BRIO consistently outperforms all baselines; on ABIDE with CC200 rather than AAL, it outperforms all baselines on both metrics despite varying baseline rankings.Source: §4.1 · p. 8

4.2 ABLATION STUDY

  • Removing re-estimation support, task relevance, or qualification-based strength modulation reduces performance, supporting their complementary roles in factor weighting and alignment.The ablations separately remove RS, TR, SM, and SF.Source: §4.2 · pp. 8–9, Figure 2 · p. 9

4.3 SENSITIVITY STUDY

  • Increasing re-estimates K generally improves AUC before gains become inconsistent, while intermediate factor counts J outperform very small or large values.The proposed explanations are diminishing benefits from larger K and a trade-off between separating patterns and preserving each factor’s capacity.Source: §4.3 · p. 9, Figure 2 · p. 9

4.4 CASE STUDY

  • Lower source-site discrepancy does not consistently correspond to lower predictive variation under FC re-estimation, so cross-site agreement alone does not imply stable within-subject predictions.At comparable discrepancy, full BRIO has lower predictive re-estimation variation than w/o QG; stronger alignment reduces both quantities in the controlled sweep.Source: §4.4 · p. 9, Figure 3 · p. 9
  • Full qualification yields the highest AUC when only some factors are retained, while its advantage narrows and disappears when all factors are retained.It outperforms task relevance alone at partial retention; support alone performs less well, indicating that both criteria matter jointly.Source: §4.4 · pp. 9–10, Figure 3 · p. 9

5 CONCLUSION

The conclusion presents BRIO as a way to account for finite-sample FC estimation variability by qualifying predictive evidence under within-scan re-estimation before cross-site alignment. Experiments across four datasets show consistent gains over competitive baselines.

5 CONCLUSION

  • BRIO combines calibrated re-estimation changes in predictive contributions with predictive relevance to determine which factors to align and how strongly.Consistently indexed connectome factors enable comparisons of contributions across full-scan and re-estimated graphs.Source: §5 · p. 10
  • Experiments on four real-world datasets show consistent gains over competitive baselines.The conclusion identifies ABIDE, REST-metaMDD, SRPBS, and ABCD as the evaluation datasets.Source: §5 · p. 10
  • The findings support estimation stability as a criterion for what to align, and motivate future study of adaptive re-estimation budgets and dynamic or multimodal connectomes.Source: §5 · p. 10

B PROOF OF PROPOSITION 1

The proof relates changes in factor contributions under FC re-estimation to class-contrast stability. Its threshold condition connects re-estimation support with whether the class contrast preserves its sign.

B PROOF OF PROPOSITION 1

  • The proof defines the class contrast from contributions averaged over subjects and FC re-estimates, then bounds its change using paired re-estimation contribution changes.The full-scan class means and their midpoint provide the reference for relating variation to support.Source: B PROOF OF PROPOSITION 1 · pp. 14–15
  • The support threshold is rewritten as a bound on re-estimation variation, using within-class variability and distance from the midpoint of the full-scan class means.Source: B PROOF OF PROPOSITION 1 · pp. 15–16
  • When the threshold holds for both classes, the re-estimated and full-scan class contrasts have the same sign; weakly separated factors require support closer to one.The threshold decreases as the separation ratio increases.Source: B PROOF OF PROPOSITION 1 · p. 16

C PROOF OF PROPOSITION 2

The proof characterizes pairwise qualification through a weighted optimization and then derives the qualification-weighted debiased Sinkhorn alignment construction. The resulting transport gradients scale by factor qualification.

C PROOF OF PROPOSITION 2

  • The qualification objective has a unique minimizer given by normalized products of task weights and factor supports, with minimum value −log g.The proof rewrites the objective as a KL divergence to that normalized weighting.Source: C PROOF OF PROPOSITION 2 · pp. 16–17
  • The support penalty is evaluated relative to task relevance, while qualification-based reweighting adjusts alignment strength according to the combined support profile.The proof relates the reduction in support penalty to reweighting from task weights.Source: C PROOF OF PROPOSITION 2 · p. 17

D DEBIASED SINKHORN DIVERGENCE

  • The debiased Sinkhorn divergence subtracts half of each self-transport cost to remove the positive bias of entropic self-transport.The construction uses the same regularization and weighted representation map for the cross- and self-transport terms.Source: D DEBIASED SINKHORN DIVERGENCE · pp. 17, 19
  • Qualification-weighted transport plans are shared across factors, while the gradient reaching each factor scales with its unnormalized qualification.The factor scales sum to the overall qualification.Source: D DEBIASED SINKHORN DIVERGENCE · pp. 18, 20

E.1 DATASET DESCRIPTION

The study uses four multi-site rs-fMRI cohorts for diagnostic classification, then applies common FC preprocessing and compares BRIO with general, graph-OOD, and brain-network baselines.

E.1 DATASET DESCRIPTION

  • ABIDE contributes 1,025 subjects from 20 sites, and REST-meta-MDD contributes 2,428 participants from 25 sites.The ABIDE cohort includes 488 individuals with ASD and 537 typically developing controls; REST-meta-MDD includes 1,300 individuals with MDD and 1,128 healthy controls.Source: §E.1 · p. 21
  • The SRPBS analysis set contains 1,000 subjects from 11 environments, while ABCD contains 425 subjects from 10 imaging sites.SRPBS compares MDD with healthy controls; ABCD uses baseline data for ADHD versus healthy-control classification.Source: §E.1 · p. 21

E.2 DATA PREPROCESSING

  • Across datasets, the common FC pipeline removes frames with non-finite ROI values, standardizes retained signals, and estimates Pearson correlations with Fisher-z transformation.The method uses AAL116 with P = 116 and CC200 with P = 200, and constructs a training-site-derived binary propagation support.Source: §E.2 · pp. 21–22
  • Within-scan FC re-estimates use shared circular moving-block samples across ROIs, preserving cross-ROI temporal correspondence and local dependence within blocks.Block length is estimated using training subjects only, with the same estimation and clipping rules applied across datasets.Source: §E.2 · p. 22

F BASELINES

  • The baseline set spans general GNNs, domain-generalization and graph-OOD methods, and specialized brain-network models.Examples include GCN, GAT, and GIN; CORAL and IRM; GSAT, DisC, CEPG, and DiSCO; and brain-network models such as BrainNetTF, XG-GNN, and BrainOOD.Source: F BASELINES · pp. 22–23

G ALGORITHM

H IMPLEMENTATION DETAILS

BRIO uses a signed graph encoder and shared factor assignments, with leave-one-site-out training and results reported across held-out sites.

H IMPLEMENTATION DETAILS

  • BRIO uses a two-layer signed GNN, four connectome factors, and a shared ROI assignment implemented with Sinkhorn iterations.The encoder has 64 hidden dimensions and 16-dimensional ROI identity embeddings; the factor mapping outputs 16-dimensional normalized representations.Source: H IMPLEMENTATION DETAILS · p. 23
  • Table 4 covers ID results on five dataset or parcellation settings, while Figure 4 presents four REST-meta-MDD analyses.Figure 4 panels cover cross-method relationships, alignment trajectories, mismatch rates, and factor retention.Source: Table 4 · p. 24, Figure 4 · p. 24
  • Each leave-one-site-out fold reserves one site for testing and two for validation, and results are reported as mean ± standard deviation across test sites.Training and training-statistic estimation use only the remaining sites.Source: H IMPLEMENTATION DETAILS · p. 24

I COMPLEXITY ANALYSIS

The paper analyzes BRIO’s amortized training cost and evaluates its ID performance using site- and class-stratified cross-validation.

I COMPLEXITY ANALYSIS

  • After supervised warm-up, per-step complexity combines graph encoding, batch-level transport, and amortized qualification refresh costs.The refresh overhead is confined to 1/M terms, and peak memory is O(BL(Pd + s) + B^2), independent of Nsrc and K.Source: I COMPLEXITY ANALYSIS · p. 24

J.1 ID PERFORMANCE COMPARISON

  • ID evaluation uses ten-fold cross-validation stratified by site and class, with a 7:2:1 train-validation-test rotation that tests each subject once.All sites appear in training, where each training site contains both classes.Source: §J.1 · p. 24
  • BRIO ranks first or second on both ID metrics across five settings, with the highest mean AUC on four and the highest mean ACC on two.BrainNetTF and DeCI lead ACC on ABIDE and ABIDE (CC200), respectively; CORE leads both metrics on SRPBS, with BRIO close behind.Source: §J.1 · p. 25

J.2 MORE SITE DISCREPANCY AND RE-ESTIMATION VARIATION

The supplementary analyses find that site discrepancy alone does not track re-estimation variation consistently, while qualification-guided factor selection and alignment outperform relevant alternatives.

J.2 MORE SITE DISCREPANCY AND RE-ESTIMATION VARIATION

  • The output-level variation metric normalizes re-estimation change against within-site, within-class full-scan prediction spread and class separation.Factor-level changes can add or cancel at the output, so the paper reports output- and factor-level stability separately.Source: §J.2 · p. 26
  • Lower site discrepancy does not consistently imply lower predictive variation under FC re-estimation across REST-meta-MDD, SRPBS, and ABCD.The metrics have positive rank correlations on REST-meta-MDD and SRPBS but zero on ABCD, with nonzero discordance on every dataset.Source: §J.2 · p. 26
  • At comparable site discrepancy, full BRIO has lower predictive variation and lower total and AUC-preserved mismatch rates than evaluated w/o QG settings on all three datasets.The controlled sweeps indicate qualification guidance benefits beyond simply increasing alignment strength.Source: §J.2 · pp. 26–27

J.3 MORE QUALIFICATION AND FACTOR TRANSFERABILITY

  • Full qualification outperforms other factor rankings at 25% or 50% retention on REST-meta-MDD, SRPBS, and ABCD.Support alone falls below random ordering on ABCD; performance gaps disappear when all factors are retained.Source: §J.3 · p. 27

J.4 MORE ABLATION STUDIES

  • On ABIDE (CC200), SRPBS, and ABCD, BRIO achieves the highest AUC and ACC among the ablation variants.The largest drop comes from removing re-estimation support on ABIDE (CC200), while removing task relevance yields the lowest performance on SRPBS and ABCD.Source: Figure 7 · p. 26, §J.4 · p. 27

J.5 MORE SENSITIVITY ANALYSIS

BRIO retains the smallest performance decline when target FC uses less scan data, while sensitivity analyses identify limits to gains from more re-estimates and show a training-resource trade-off.

J.5 MORE SENSITIVITY ANALYSIS

  • Increasing K generally improves AUC before gains become inconsistent, while intermediate J values tend to outperform very small or large factor counts.The analysis extends to ABIDE (CC200), SRPBS, and ABCD.Source: §J.5 · p. 27
  • The default temporal block length achieves the highest mean AUC and ACC in all five settings; halving or doubling it degrades performance.Halving is more harmful on most datasets; estimated defaults correspond to similar physical durations across datasets despite different repetition times.Source: §J.5 · p. 27

J.6 EFFECT OF SCAN LENGTH ON PREDICTIVE PERFORMANCE

  • BRIO shows the smallest AUC decline when target FC uses fewer time points, compared with BRIO w/o RS and BRIO w/o QG.The decline is larger at 50% than at 75% of the preprocessed BOLD sequence; Fig. 9 reports the reduced-scan comparison.Source: §J.6 · p. 28, Figure 9 · p. 28

J.7 TRAINING EFFICIENCY AND RESOURCE CONSUMPTION

  • Across four datasets, BRIO uses more time and memory than FC-HGNN and CORE, less time but slightly more memory than BrainOOD, and less of both than DiSCO.Table 6 reports per-epoch training time and GPU memory consumption.Source: §J.7 · p. 28, Table 6 · p. 28

K INTERPRETABILITY ANALYSIS

BRIO projects qualification-weighted factor emphasis into ROI profiles and aggregates those profiles across folds and seeds. The selected regions overlap partly between ABIDE and REST-meta-MDD.

K INTERPRETABILITY ANALYSIS

  • Each ROI’s normalized profile score reflects pooling emphasis in factors with higher average source qualification; uniform factor weights yield equal scores across ROIs.The profile is formed by projecting qualifications through the retained factor–ROI assignment.Source: K INTERPRETABILITY ANALYSIS · p. 29
  • Profiles are averaged across evaluated folds and seeds in the common AAL116 ROI space because factor indices may not correspond across independently trained models.This preserves each model’s factor–ROI correspondence during projection and yields a dataset-level summary.Source: K INTERPRETABILITY ANALYSIS · p. 29
  • The right posterior cingulate and left medial superior frontal regions rank among the five highest-scoring ROIs in both ABIDE and REST-meta-MDD, while the other selected ROIs differ.ABIDE’s list also includes the left precuneus, right superior temporal gyrus, and right fusiform gyrus; REST-meta-MDD’s includes the left amygdala, right hippocampus, and right anterior cingulate gyrus.Source: K INTERPRETABILITY ANALYSIS · p. 29, Figure 10 · p. 29
Loading 2609.34611v1…