acceptodds
Under review as a conference paper at ICLR 2027

AREION: A Foundation Model for Atlas-Agnostic Representation Learning of Brain Connectome

Abstract

Brain foundation models have shown promising advances in learning generalizable representations from functional magnetic resonance imaging (fMRI) data. However, existing models are typically tied to specific brain atlases and cannot be applied to fMRI data represented using different atlases without full retraining, thereby severely limiting their broader utility across neuroimaging studies. To address this bottleneck, we introduce AREION, atlas-agnostic representation learning of brain connectomes. AREION overcomes atlas dependency through two integrated innovations. First, a Brain Region Contextualizing technique maps brain regions and connectivity entries from diverse atlases into a unified representation space based on their spatial, functional, and morphological characteristics. Second, a Hierarchical Brain Transformer Encoder processes region-specific connectivity profiles as variable-length token sequences and jointly captures local functional structure and global brain functional organization, conditioned on brain regional context, without assuming a fixed number of regions. Together, these two techniques enable the learning of generalizable representations across brain atlases with different parcellation resolutions and connectivity dimensions. AREION is pretrained on a large-scale multi-atlas cohort of fMRI data via self-supervised modeling and evaluated across diverse downstream tasks. Extensive experiments show that it outperforms strong supervised baselines and fMRI foundation models while using substantially fewer parameters, demonstrating strong transferability to atlases unseen during pretraining, generalization to external datasets, and excellent interpretability at the brain regional level.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Brain foundation model, Brain connectome, fMRI

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_8rtAOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents AREION, a foundation model for atlas-agnostic representation learning from fMRI functional connectivity (FC) matrices. The method combines Brain Region Contextualizing (BRC), which incorporates spatial coordinates, functional gradients, and regional volume, with a hierarchical Transformer that models regional connectivity profiles and whole-brain interactions. The model is pretrained using masked autoencoding on multi-atlas fMRI data and evaluated on several downstream tasks, including demographic, cognitive, and disease-related prediction, as well as generalization to unseen atlases and external datasets.

Critical strengths
  1. The paper addresses the practical challenge of learning transferable representations across different brain atlases.
  2. The combination of region-level metadata and hierarchical Transformer encoding is reasonable.
  3. The paper is easy to follow.
Critical weaknesses
  1. The single-atlas and multi-atlas settings differ not only in atlas diversity but also in the number of training subjects and dataset composition. In particular, Gordon333 is used only for ABCD, while the Schaefer atlases are used for HCP/HCP-D. Therefore, the reported improvements cannot be clearly attributed to atlas diversity rather than additional data or cohort differences.
  2. Functional gradients are an important component of BRC, but the paper does not sufficiently explain how they are computed, which reference data are used, or how they are aligned and normalized across different atlases.
  3. Since FC matrices are symmetric ((F_{ij}=F_{ji})), masking only the rows of selected regions while retaining all columns in visible rows would expose some reconstruction targets through their symmetric entries. With a 75% masking ratio, approximately 25% of the entries in each masked row could remain observable. The paper does not clearly specify whether the corresponding columns are also masked.
  4. Some reported improvements over baselines are relatively small compared with the reported cross-validation variability. Although the paper provides results from five-fold cross-validation, it does not report paired significance tests or confidence intervals for the performance differences.
  5. Table 4 lists SCZ with Glasser360 as an external test setting for schizophrenia classification, but the corresponding results are not reported in the main paper or appendix.
Questions for authors
  1. Can the authors provide a controlled comparison between single-atlas and multi-atlas pretraining using matched training subjects and a comparable training budget?
  2. How are the functional gradients computed, aligned, and normalized across different atlases, particularly those unseen during pretraining?
  3. During masked autoencoding, are the connectivity entries corresponding to masked regions also removed from the visible rows? If not, how is information exposure through the symmetry of the FC matrix addressed?
  4. Can the authors provide paired statistical significance tests or confidence intervals for the main performance comparisons?
  5. Can the authors provide the SCZ–Glasser360 schizophrenia classification results listed as an external evaluation setting in Table 4?
1 comment

We thank the reviewer for the insightful assessment of our work. All responses to the identified weaknesses can be found within our answers to the reviewer’s questions, as each weakness directly corresponds to a specific question. Therefore, detailed responses addressing all weaknesses are provided in the following question-by-question replies.

Q1: Our results indicate that the performance improvement comes from both the proposed regional contextualizing mechanism and multi-atlas pretraining. In Table 1, AREION (Single), pretrained only on Schaefer400, already outperforms all baselines on five of the six tasks, showing the effectiveness of our framework without multi-atlas pretraining. In Table 3, in which all variants share identical pretraining data, BRC outperforms alternative positional encodings, particularly on the unseen Gordon333 atlas (91.61% vs. 88.81% ACC for sinusoidal PE).

Q2: The voxel-wise functional gradient maps from the HCP cohort were used in BRC, publicly available at https://github.com/PennLINC/Brain_Organization/tree/master/PrincipleGradient. For each atlas, region-wise mean and standard deviation of the first three principal functional gradients were included in BRC. We did not compute functional gradients for each subject. Given a new/unseen atlas, its BRC could be computed directly with these group-level functional gradient maps.

Q3: In our implementation, the columns corresponding to masked regions are retained in the visible rows. The reasons are as follows: 1. Whether the columns corresponding to the masked regions are removed only affects the difficulty of pretraining task. Our experiments demonstrate that the representations learned under this setting achieve strong performance across downstream tasks. 2. Retaining the columns preserves whole-brain context. With a 75% masking ratio, masking both rows and columns would leave each visible region with connectivity to only 25% of the regions, i.e., only 6.25% of the FC matrix would be observed. The Regional Transformer Encoder would then learn from truncated connectivity profiles, whereas at inference it receives the complete profile of each region. While symmetric masking provides a good future direction to go, we anticipate that adjusting mask ratio may be needed for different masking strategies.

Q4: We performed subject-level paired comparisons between AREION and the strongest baseline for each task in Table 1, using the same subject-level cross-validation splits. We report the pooled out-of-fold performance, the paired 95% bootstrap confidence interval of the difference and the two-sided Wilcoxon signed-rank test p-value computed on per-subject errors.

TaskStrongest baselineMetricAREIONBaseline95% Paired CIpwilcoxonp_{\text{wilcoxon}}
HCP (Sex)BrainNetTFACC92.85%92.12%[-1.04%, 2.59%]0.437
SCZ (SCHZ)SVMACC72.48%71.56%[-3.21%, 5.50%]0.683
HCP (Age)BrainNetCNNR2R^20.3020.262[0.010, 0.069]0.047
PNC (Age)RidgeR2R^20.6430.633[-0.004, 0.026]0.690
PNC (Exec_Func)BrainNetTFR2R^20.3880.385[-0.026, 0.033]0.944

AREION improves over the strongest baseline on all five tasks, and the improvement is significant on HCP Age. To assess overall improvement on a common scale, we computed each task's relative gain, (AREION−Baseline)/Baseline(\text{AREION}-\text{Baseline})/\text{Baseline} and took the unweighted average across the five tasks using 10,000 cohort-stratified subject-level paired bootstrap resamples—resampling subjects with replacement within each of the three cohorts (HCP, SCZ, PNC) to preserve shared-subject correlations, recomputing each task's metric and relative gain, and averaging the five relative gains per replicate. This yields a mean relative improvement of +4.07% (95% CI [+0.96%, +7.65%], two-sided p = 0.009).

Q5: We provide the SCZ-Glasser360 schizophrenia classification results as below:

MethodACC↑AUC↑F1↑
SVM71.99 ± 3.3179.14 ± 4.1971.25 ± 3.75
BrainGFM69.77 ± 8.2277.94 ± 6.6469.77 ± 8.32
AREION72.45 ± 3.7182.08 ± 3.1871.42 ± 6.09

This result further demonstrates the strong transferability to external dataset and unseen atlas of our proposed framework.

Reviewer_Tr5EOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces AREION, a foundation model for atlas-agnostic representation learning of functional brain connectomes. The method addresses the difficulty of transferring existing brain foundation models across different brain parcellation schemes. It introduces Brain Region Contextualizing to encode spatial, functional, and morphological properties of brain regions, together with a hierarchical Transformer that processes variable-length connectivity representations. The model is pretrained using functional connectivity data from multiple brain atlases and evaluated on several downstream prediction tasks, including classification and regression. The experiments further investigate generalization to previously unseen brain atlases.

Critical strengths
  1. The paper addresses an important limitation of existing brain foundation models. Different neuroimaging studies frequently adopt different brain atlases, and the resulting variation in connectivity dimensions makes cross-dataset transfer challenging. The proposed atlas-agnostic formulation is well motivated and has practical significance.

  2. The combination of Brain Region Contextualizing and hierarchical connectivity modeling provides a coherent solution to the problem. Encoding regional properties rather than relying only on atlas identifiers is a meaningful design choice, particularly for transferring representations to unseen parcellation schemes.

  3. The experimental evaluation is relatively comprehensive. The paper considers multiple downstream tasks, compares against conventional machine learning approaches and existing foundation models, and explicitly evaluates transfer to unseen atlases. The results provide encouraging evidence that the proposed representation learning strategy improves generalization.

Critical weaknesses
  1. Although the overall framework is well motivated, the individual architectural components are largely based on established representation learning techniques. The main contribution lies in adapting and integrating these techniques for atlas-agnostic brain modeling. The paper could more clearly distinguish its methodological novelty from previous approaches using spatial embeddings and hierarchical attention.

  2. The contribution of individual regional properties requires further investigation. Spatial coordinates, functional gradients, and morphological characteristics may contain overlapping information. While the paper includes component ablations, a deeper analysis of how these properties influence cross-atlas transfer would improve the interpretation of the results.

  3. The generalization experiments are encouraging, but the evaluation could be strengthened by considering larger differences in acquisition protocols, preprocessing pipelines, and population characteristics. Generalization across atlases does not necessarily imply robustness to all forms of neuroimaging domain shift.

Questions for authors
  1. How does the model perform when only spatial coordinates are available for an unseen atlas, without reliable functional or morphological descriptors?

  2. To what extent does the performance improvement come from multi-atlas pretraining rather than the proposed regional contextualization mechanism?

  3. Could the authors provide additional analysis of the computational cost as the number of brain regions increases?

2 comments

Response to Reviewer Tr5E (1/2): W1–W3

We thank the reviewer for the insightful assessment and address concerns as follows:

W1: Our major contribution lies in the representation of brain regions and connectomes, which enables a single model to operate across arbitrary atlases. Previous methods that use spatial or functional embeddings, such as the gradient positioning in Brain-JEPA, represent brain regions from a predefined atlas. Consequently, they cannot be directly applied to unseen parcellations with different regional definitions and region-wise FC profiles of different dimensionalities. In contrast, BRC characterizes each region according to its spatial, functional and morphological properties. Because these properties can be computed for any atlases, enabling transfer to atlases unseen during pretraining without retraining.

Additionally, the hierarchical encoder further introduces a new approach to tokenizing brain connectomes. Previous Transformer-based models typically represent the entire connectivity profile of each region as one token, making the token dimension dependent on the number of regions in the atlas. AREION instead represents each pairwise connection as a token and hierarchically aggregates connection-level information into regional and whole-brain representations. This design allows the same model architecture and parameter set to process atlases with any number of regions.

W2: The three properties are derived from different sources and characterize different aspects of a region. Coordinates describe its location, functional gradients are derived from functional connectivity, and volume reflects the resolution of the parcellation. The ablation study results shown in Figure 3(a), with numerical results in our response to Q1, supports this. Adding gradients to coordinates improves performance in all settings, and adding volume brings more improvement in most of them. For cross-atlas transfer, the benefit of gradients is much larger on unseen atlases, for example an improvement by 2.16 points on PNC with the unseen Glasser360 atlas compared with 0.14 points with Schaefer400, suggesting that gradients help the model relate regions across parcellations whose boundaries differ.

W3: Thank you for the insightful comment. We agree that generalization across atlases does not imply robustness to every form of domain shift. Our external evaluation already involves shifts in all three aspects the reviewer mentions. PNC and SCZ were collected in independent studies that are not part of the pretraining data, and both were preprocessed with pipelines different from those of the pretraining cohorts, following (Cui et al., 2020) for PNC and fMRIPrep with ICA-AROMA for SCZ. They also differ in demographic and clinical characteristics: PNC is a community sample of children and adolescents aged 8 to 23, and SCZ is a clinical cohort of 117 healthy controls and 101 patients with schizophrenia. AREION performs well on both datasets, including when they are represented with atlases unseen during pretraining (Tables 1 and 2). Evaluating robustness to an even wider range of acquisition protocols and populations is an important direction for future work, and we expect that pre-training with data from larger, diverse cohort could potentially enhance the robustness of the pretrained model.

Response to Reviewer Tr5E (2/2): Q1–Q3

Q1: Thank you for the insightful comment. This setting is evaluated in our BRC ablation (Section 4.3.2, Figure 3(a)), in which the “coordinates” variant uses only the spatial coordinates as the regional context. Importantly, all ablation variants are pretrained exclusively on the HCP dataset with the Schaefer100/200/300/400 atlases. Therefore, Gordon333 and Glasser360 are both unseen during pretraining, and PNC serves as an external dataset. The four evaluation settings thus cover: (i) a seen dataset with a seen atlas (HCP, Schaefer400), (ii) a seen dataset with an unseen atlas (HCP, Gordon333), (iii) an external dataset with a seen atlas (PNC, Schaefer400), and (iv) an external dataset with an unseen atlas (PNC, Glasser360). For clarity, we report the numerical results below (sex classification, ACC, %):

BRC inputsHCP (Schaefer400)HCP (Gordon333)PNC (Schaefer400)PNC (Glasser360)
coordinates only92.4490.7884.5683.84
coordinates+gradients92.7791.4084.7086.00
Full BRC92.9691.6185.1386.00

Furthermore, we used publicly available, group-level functional gradient maps in BRC, so that it could be applied to new atlas directly if the voxel-wise parcellation assignment is available.

Q2: Our results indicate that the performance improvement comes from both the proposed regional contextualizing mechanism and multi-atlas pretraining. In Table 1, AREION (Single), pretrained only on Schaefer400, already outperforms all baselines on five of the six tasks, showing the effectiveness of our framework without multi-atlas pretraining. In Table 3, where all variants share identical pretraining data, BRC outperforms alternative positional encodings, particularly on the unseen Gordon333 atlas (91.61% vs. 88.81% ACC for sinusoidal PE).

Q3: We measured the cost of the AREION encoder on a single NVIDIA B200 GPU for atlases with 100/200/300/400 regions, both for inference with the frozen encoder and one pretraining step with 75% masking. The results are shown in the table below.

#ROIInference time per subjectInference GFLOPs per subjectInference memory (batch_size=8)Pretraining time per samplePretraining GFLOPs per samplePretraining memory (batch_size=8)
1001.5 ms47.30.78 GB3.9 ms37.51.54 GB
2006.1 ms203.33.00 GB8.9 ms159.95.36 GB
30013.6 ms492.66.70 GB15.8 ms388.811.64 GB
40026.0 ms939.911.87 GB28.0 ms745.620.26 GB

As the number of regions increases, the number of parameters stays fixed at 6.4M, since no layer of AREION is sized by the number of regions. Time, computation and memory increase steadily with the number of regions and remain modest at 400 regions, where inference takes 26 ms per subject and pretraining 28 ms per sample.

Reviewer_M4keOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AREION, an atlas-agnostic foundation model for brain functional connectivity representation learning. The main contribution is Brain Region Contextualizing (BRC), which incorporates spatial coordinates, functional gradients, and regional volume to characterize brain regions across different atlases. The paper addresses an important research direction toward atlas-agnostic brain foundation models. However, several key design choices are insufficiently motivated or supported by appropriate ablation studies. Moreover, the evaluation protocols, particularly those concerning cross-atlas transfer and downstream adaptation, are not sufficiently justified or investigated. These issues make it difficult to assess whether the proposed method fully achieves its stated objectives.

Critical strengths
  1. The paper addresses atlas dependency in brain foundation models, a practical and important limitation that affects their generalizability across neuroimaging datasets.
  2. Incorporating spatial, functional, and morphological information into regional representations is a promising direction.
Critical weaknesses
  1. Although AREION claims to learn atlas-agnostic representations, the final flattened features remain atlas-dependent in dimensionality, requiring separate predictors for different atlases. The authors acknowledge in Lines 315–317 that representations from different atlases could be integrated using Transformers, but do not justify their choice of atlas-specific predictors. Furthermore, flattening regional embeddings produces extremely high-dimensional features (e.g., 102,400 dimensions for Schaefer400), potentially confounding comparisons with baseline foundation models. Dimension-matched comparisons and shared-predictor evaluations are needed.

  2. The 13-dimensional BRC design is insufficiently justified. And the paper does not explain how functional gradients are obtained for different subjects and atlases. Since functional gradients are central to BRC, their computation and sensitivity to different reference choices should be clarified and evaluated.

  3. Most downstream experiments involve datasets already used during pretraining. More independent external datasets are needed to establish robust generalizability.

  4. The paper does not discuss Toward a Multi-View Brain Network Foundation Model: Cross-View Consistency Learning Across Arbitrary Atlases, which shares a closely related motivation and solution, and explores cross-atlas representation learning.

  5. The choice of atlases is insufficiently justified.

  6. Voxel-level fMRI foundation models, which do not require predefined atlas parcellations, are another relevant direction toward atlas-agnostic learning and should be discussed.

Questions for authors
  1. In Lines 315–317, the authors state that representations from atlases with different numbers of regions can be integrated using Transformers. Why does the paper nevertheless adopt atlas-specific flattened representations and separate traditional machine learning predictors? What prevents the authors from evaluating a shared downstream predictor across different atlases?

  2. How exactly are the functional gradients computed?

  3. What motivates the specific 13-dimensional design of BRC? Any ablations?

2 comments

Response to Reviewer M4ke (1/2): W1–W4, Q1

We thank the reviewer for the constructive assessment and address concerns as follows.

W1 & Q1: We would like to clarify that the atlas-agnostic representation learning of AREION enables a single pretrained model to learn from FC matrices derived using any atlas, including atlases not seen during pretraining, while embedding region tokens from all atlases in a shared representation space. Atlas dependence arises only at the downstream application stage when the region tokens are flattened into a single vector, whose dimension varies with the number of regions.

Following the reviewer’s suggestion, we extended our proposed framework by introducing learnable brain tokens into the Brain-wide Transformer Encoder. Through self-attention, these brain tokens aggregate information from each FC matrix into a representation of fixed dimension, regardless of the atlas used (Appendix B.3). A unified predictor trained on the fixed-dimension brain tokens across atlases (Schaefer100/200/300/400) generally outperforms atlas-specific predictors. This result demonstrates both the generalizability of the brain-token representation across atlases and the benefit of jointly leveraging representations learned from multiple brain atlases.

Regarding the feature dimension, the brain-token representation of Schaefer400 has less than one third the dimension of the flattened region token representation and is also smaller than the raw FC feature vector. Despite this, AREION continues to outperform most baselines in Table 1, including the evaluated foundation models. These results indicate that AREION’s advantage is not attributable to a larger feature dimension.

W2: We address this concern by answering Q2 and Q3 below.

W3: We thank the reviewer for this comment and would like to clarify that most of our downstream evaluations were conducted on external datasets that are not seen during pretraining. As listed in Table 4, PNC and SCZ are external datasets. Four of the six tasks in Table 1, namely PNC Sex, PNC Age, PNC Exec_Func and SCZ, were evaluated on these two external datasets, and all experiments in Table 2 using PNC with atlases that were also not seen during pretraining. To further address this concern, we report results on SCZ with the Glasser360 atlas, with both the dataset and atlas not seen during pretraining.

MethodACC↑AUC↑F1↑
SVM71.99 ± 3.3179.14 ± 4.1971.25 ± 3.75
BrainGFM69.77 ± 8.2277.94 ± 6.6469.77 ± 8.32
AREION72.45 ± 3.7182.08 ± 3.1871.42 ± 6.09

W4: We have added a discussion of MV-BrainFM to the introduction and related work sections of the revised manuscript. MV-BrainFM aligns representations derived from different atlas views of the same participant through cross-view consistency learning and incorporates spatial priors using coordinate features. However, adaption to an unseen atlas requires a new atlas-specific input projection, which is then fine-tuned together with the backbone and prediction head.

In contrast, AREION represents each functional connectivity value as a token conditioned on the spatial, functional, and morphological properties of the corresponding pair of regions. It employs a shared input projection that is independent on the number of regions in the atlas. This design enables the pretrained AREION encoder to process previously unseen atlases without any atlas-specific parameters or retraining the encoder. At the same time, we find that the cross-view/atlas consistency learning technique in MV-BrainFM is insightful and incorporating a related cross-atlas consistency objective may further enhance AREION’s ability to learn fixed-dimension brain-wide representation that are shared across atlases.

Response to Reviewer M4ke (2/2): W5–W6, Q2–Q3

W5: The rationale for selecting these atlases is provided in Appendix D and Table 9. The atlases were chosen to capture two complementary sources of variation across parcellations: spatial resolution and the principles used to define regions. The Schaefer atlases (Schaefer100/200/300/400) are among the most widely used functional parcellations, and their four resolutions provide progressively finer granularity while preserving a consistent functional organization, allowing us to assess robustness to variation in the number and granularity of brain regions. Moreover, the range of 100 to 400 regions encompasses the spatial resolutions most commonly used for brain connectome analyses.

Gordon333 is a functional atlas developed independently of the Schaefer series and therefore introduces differences in regional boundaries and functional organization beyond those attributable to resolution alone. Glasser360 and the Brainnetome Atlas were excluded from pretraining and used specifically to evaluate generalization to unseen atlases. Glasser360 is a multimodal cortical parcellation, whereas the Brainnetome Atlas is built from connectivity patterns in multimodal neuroimaging data. Thus, both differ from the pretraining atlases in how their regions are defined and provide complementary tests of AREION’s ability to generalize to previously unseen atlases.

W6: We thank the reviewer for the insightful comment. We have added a discussion of voxel-level fMRI foundation models in the section of related works (Section 2.3) of the revised manuscript.

Q2: The voxel-wise functional gradient maps from the HCP cohort were used in BRC, publicly available at https://github.com/PennLINC/Brain_Organization/tree/master/PrincipleGradient. For each atlas, region-wise mean and standard deviation of the first three principal functional gradients were included in BRC. We did not compute individualized functional gradients for each subject. Given a new/unseen atlas, its BRC could be computed directly with these group-level functional gradient maps.

Q3: The BRC mechanism is motivated by the need to characterize each brain region according to its intrinsic properties, rather than an atlas-specific index, thereby enabling a single encoder to process arbitrary atlases, including those unseen during pretraining. BRC characterizes each region from three complementary perspectives: spatial coordinates, functional gradients, and regional volume.

We conducted two sets of ablation experiments to evaluate BRC. First, we compared BRC with commonly used positional encodings (PE). As shown in Table 3, BRC outperforms sinusoidal PE, relative PE, and the variant without PE under identical pretraining settings. The largest improvement is observed on the unseen Gordon333 atlas (91.61% vs. 88.81% ACC for sinusoidal PE).

Second, we evaluated the individual contributions of the three components of BRC. As shown in Figure 3(a), starting with spatial coordinates and progressively adding functional gradients and regional volume consistently improves performance on both seen and unseen atlases. These results indicate that the three components capture complementary information and collectively contribute to BRC’s effectiveness.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.