CoFlu-Brain: An fMRI Connectivity Foundation Model with Cofluctuation-Based Pretraining

Ziteng Sui; Junfeng Xia; Wenhao Ye; Mo Wang; Quanying Liu

cs.NEOct 11, 2026

CoFlu-Brain: An fMRI Connectivity Foundation Model with Cofluctuation-Based Pretraining

Authors: Ziteng Sui, Junfeng Xia, Wenhao Ye, Mo Wang, Quanying Liu

Organizations: Department of Biomedical Engineering, Southern University of Science and Technology, China · School of Biomedical Engineering, Shenzhen University, China · Department of Computer Science, University of Warwick, UK

Abstract

Functional magnetic resonance imaging (fMRI) commonly characterizes large-scale brain organization through functional connectivity (FC). However, an observed FC matrix is not a fixed brain graph but a statistical estimate determined by the finite BOLD frames used to compute it. Conventional FC-based approaches typically represent each scan with a single Pearson-correlation matrix, leaving frame-resolved variations in regional interactions unexploited. To address this limitation, we introduce CoFlu-Brain, which organizes framewise cofluctuation observations into multiple retained and dropped views, each summarized as an FC matrix. The model learns connectivity structure reproducible across retained views and predicts the representations of the corresponding dropped views, while adopting SIGReg from LeJEPA to prevent representation collapse. We pretrained CoFlu-Brain on 13,444 subjects from eight dataset groups and evaluated its representations via frozen-encoder linear probing on seven classification and two age-prediction tasks across seven datasets. CoFlu-Brain showed strong transfer across demographic, developmental, and clinical prediction tasks, outperforming the evaluated released models on most tasks. Random frame partitioning improved aggregate transfer, while the complete objective achieved the strongest overall performance. Code is available.

Figures & tables

Figure 1: From a single FC estimate to cofluctuation based multi-view learning. Conventional FC aggregates all frames into a single connectivity estimate, whereas CoFlu-Brain uses different compositions of framewise cofluctuation observations to construct multiple FC views.

Figure 2: CoFlu-Brain pretraining and downstream evaluation. (a) Framewise cofluctuations from standardized regional BOLD signals. (b) Paired retained–dropped FC views train a shared encoder through consistency, latent prediction, and SIGReg. (c) Frozen-encoder linear probing and downstream inference.

Panel of Table 1 — Table 1: Linear probing of fMRI models for classification and age prediction. Values are mean $\pm$ sample SD across five participant-grouped folds. Classification tasks report accuracy and weighted F1 ( $\uparrow$ ); age-prediction tasks report nMAE ( $\downarrow$ ) and Pearson’s $r$ ( $\uparrow$ ). Red boldface marks the best mean for each metric. CoFlu-Brain averages three pretrained checkpoints per fold; baselines use one released checkpoint. Overall rank test: cohort-blocked permutation $p<0.001$ .

Figure 3: Network-restricted probing on ADHD200. (a) Age nMAE ( $\downarrow$ ). (b) Diagnosis weighted F1 ( $\uparrow$ ). Cortical maps summarize mean performance across all 17 networks, and the corresponding point-range plots detail the five best-performing networks for each task. Points indicate individual pretraining seeds, diamonds indicate seed means, horizontal segments indicate seed ranges, and dashed vertical lines indicate whole-brain performance.

Figure 4: Random versus ordered frame partitioning. (a) Relative weighted-F1 gain for classification and relative nMAE reduction for age prediction; positive values favor random partitioning. (b) Mean weighted F1 across seven classification tasks. (c) Mean nMAE across two age-prediction tasks. Lines connect paired pretraining seeds; red diamonds denote seed means.

Cohort-averaged rank: Full 1.50; No prediction 3.57; No consistency 1.79; No SIGReg 3.14
Dataset Objective	HCP / Sex	ADNI / AD	ADHD200 / Diagnosis	ADHD200 / Age
	Weighted F1 $\uparrow$	Weighted F1 $\uparrow$	Weighted F1 $\uparrow$	nMAE $\downarrow$
Full	0.7874 $\pm$ 0.0115	0.7219 $\pm$ 0.0657	0.5872 $\pm$ 0.0215	0.6523 $\pm$ 0.0315
No prediction	0.6565 $\pm$ 0.0187	0.5023 $\pm$ 0.0850	0.5348 $\pm$ 0.0309	0.7416 $\pm$ 0.0567
No consistency	0.7789 $\pm$ 0.0110	0.6866 $\pm$ 0.1175	0.5725 $\pm$ 0.0384	0.6611 $\pm$ 0.0539
No SIGReg	0.6942 $\pm$ 0.0153	0.6065 $\pm$ 0.0667	0.5643 $\pm$ 0.0288	0.7132 $\pm$ 0.0482

Table 2: Downstream performance under objective-component ablation. Each objective variant was pretrained independently with seed 42 under matched configurations and is reported as mean

\pm

SD across five participant-grouped folds. Metrics are weighted F1 (

\uparrow

) and age nMAE (

\downarrow

). Red boldface marks the best mean for each task. Overall rank test: cohort-blocked permutation

p=0.0014

Panel of Table 3 — Table 3: Pretraining data coverage and downstream performance. Values are mean $\pm$ SD across probe folds. Metrics are weighted F1 ( $\uparrow$ ) and age nMAE ( $\downarrow$ ). Red boldface marks the best mean for each task.

Panel of Table 4 — Table 4: Downstream performance by temporal branch count. Values are mean $\pm$ SD across probe folds. Metrics are weighted F1 ( $\uparrow$ ) and age nMAE ( $\downarrow$ ). Red boldface marks the best mean for each task.

Panel of Table 5 — Table 5: Downstream performance under consistency-weight schedules. Values are mean $\pm$ SD across probe folds. Metrics are weighted F1 ( $\uparrow$ ) and age nMAE ( $\downarrow$ ). Red boldface marks the best mean for each task.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset / task	Participants	Examples	Target
HCP sex	1,000	5,997	Binary sex
ABCD sex	500	999	Binary sex
PPMI group	474	1,896	Three diagnostic groups
ADNI AD	230	230	AD vs. control
ADNI MCI	292	292	MCI vs. control
ABIDE ASD	173	173	ASD vs. control

Appendix

Table 6: Downstream evaluation cohorts and prediction targets. Participants are grouped across all available segment-level examples when constructing cross-validation folds.

Model configuration
Configuration	Value
Input representation
ROI atlas	Schaefer-400
Input segment length	200 frames
FC-row embedder
MLP dimensions	$400\rightarrow 512\rightarrow 128\rightarrow 512$

Appendix

Table 7: CoFlu-Brain model, pretraining, and linear-probe configurations.

Panel of Table 8 — Table 8: CoFlu-Brain linear probing across pretraining seeds. Values are mean $\pm$ sample SD across five participant-grouped folds. Metrics are weighted F1 ( $\uparrow$ ) and age nMAE ( $\downarrow$ ). The final column averages scores across the three pretraining seeds within each fold before computing mean $\pm$ SD.

Network	ROIs	Age nMAE $\downarrow$	Diagnosis F1 $\uparrow$
DefaultA	34	0.6804 $\pm$ 0.0065	0.5965 $\pm$ 0.0165
ContC	12	0.6894 $\pm$ 0.0029	0.5619 $\pm$ 0.0132
LimbicA	13	0.6923 $\pm$ 0.0056	0.5442 $\pm$ 0.0047
DefaultC	13	0.7012 $\pm$ 0.0052	0.5288 $\pm$ 0.0125
SomMotA	39	0.7042 $\pm$ 0.0116	0.5435 $\pm$ 0.0226
LimbicB	11	0.7196 $\pm$ 0.0138	0.5372 $\pm$ 0.0092

Appendix

Table 9: Functional-network selection on ADHD200. Gray shading marks the top-five networks identified for each task by network-restricted probing. Values are mean

\pm

SD across three independent pretraining runs. Age prediction reports nMAE (

\downarrow

); diagnosis reports weighted F1 (

\uparrow

). ROI counts indicate network size.

Panel of Table 10 — Table 10: Downstream performance under random and ordered frame partitioning. Each seed entry is mean $\pm$ sample SD across five participant-grouped folds. The final column averages scores across pretraining seeds within each fold before computing mean $\pm$ SD. Metrics are weighted F1 ( $\uparrow$ ) and age nMAE ( $\downarrow$ ). Red boldface marks the better seed-averaged score for each task.

Panel of Table 11 — Table 11: Objective-component ablation across nine downstream tasks. Each objective variant was pretrained independently with seed 42 under matched configurations and is reported as mean $\pm$ SD across five participant-grouped folds. Metrics are weighted F1 ( $\uparrow$ ) and age nMAE ( $\downarrow$ ). Red boldface marks the best mean for each task.

Panel of Table 12 — Table 12: Pretraining data coverage and downstream performance across nine tasks. Values are mean $\pm$ SD across probe folds. Metrics are weighted F1 ( $\uparrow$ ) and age nMAE ( $\downarrow$ ). Red boldface marks the best mean for each task.

Panel of Table 13 — Table 13: Temporal branch-count comparison across nine downstream tasks. Values are mean $\pm$ SD across probe folds. Metrics are weighted F1 ( $\uparrow$ ) and age nMAE ( $\downarrow$ ). Red boldface marks the best mean for each task.

Panel of Table 14 — Table 14: Consistency-weight schedule comparison across nine downstream tasks. Values are mean $\pm$ SD across probe folds. Metrics are weighted F1 ( $\uparrow$ ) and age nMAE ( $\downarrow$ ). Red boldface marks the best mean for each task.

Explore similar work

Flattening the Connectome Spectrum: A Spectral Filter for FC Induces a Pretraining Target for fMRI Encoders
Sep 29, 2026Giovanni Marraffini, Victoria Shevchenko, Carlo Alberto Barbano +1Kernel Ridge Regression Medical Imaging Foundation Models
UNITO · MIND
The Variance Brain Foundation Models Forgot: Third-Order Statistics Predict Cognition Where Billion-Parameter Models Fail
May 29, 2026Giovanni Marraffini, Gabriel Mahuas, Trinidad Borrell +2Medical Imaging Foundation Models Functional Connectivity
Inria Saclay Île-de-France, CEA, Université Paris-Saclay, Palaiseau, France · Sigma Nova · Sorbonne Université, Institut du Cerveau - Paris Brain Institute - ICM +1
Scaling Vision Transformers for Functional MRI with Flat Maps
Oct 15, 2025Connor Lane, Mihir Tripathy, Leema Krishna Murali +15Medical Imaging Foundation Models Masked Autoencoders
MedARC · Sophont · Baylor College of Medicine +3

Sep 29, 2026J/K move · Enter open · S save

Flattening the Connectome Spectrum: A Spectral Filter for FC Induces a Pretraining Target for fMRI Encoders

Giovanni Marraffini, Victoria Shevchenko, Carlo Alberto Barbano, Demian Wassermann

UNITO · MIND

Kernel Ridge Regression Medical Imaging Foundation Models Functional Connectivity Self-Supervised Pre-Training fMRI Foundation Models Spectral Methods

Open paper AskPDF arXiv

Self-supervised pretraining reshaped prediction in language and vision, and brain foundation models (BFMs) inherited its promise. Representations learned from large unlabelled corpora should capture individual functional dynamics and generalise across cohorts. However, kernel ridge regression (KRR) fitted on functional connectivity (FC) matrices still predicts individual phenotypes more accurately than any BFM we tested. In this paper, we show that KRR is weighted by the eigenvalues of the FC which are miscalibrated for phenotype prediction. We apply an efficient spectral filter to recalibrate the eigenvalues of each subject's FC matrix, enabling the model to exploit more inter-individual variance. Across the 5 datasets, 11 parcellations and 6 prediction targets we tested, we match or exceed the KRR baseline. Based on this finding, we then pretrain a small encoder model on about 4,000 hours of fMRI from 162 open datasets, whereby we align the pairwise similarities between the embeddings of recording snippets with those between the recalibrated connectomes. Our model performs on par with the best of the 6 published BFMs we tested while having an order of magnitude fewer parameters. Our encoder performs better than FC on short scans and in smaller cohorts, especially in fingerprinting. We release the pretrained model weights, the code and the pretraining data, preprocessed and parcellated.

Abstract

Explore similar work

Flattening the Connectome Spectrum: A Spectral Filter for FC Induces a Pretraining Target for fMRI Encoders

The Variance Brain Foundation Models Forgot: Third-Order Statistics Predict Cognition Where Billion-Parameter Models Fail

Scaling Vision Transformers for Functional MRI with Flat Maps