Organizations: Saw Swee Hock School of Public Health at National University of Singapore, Singapore · School of Biomedical Engineering & State Key Laboratory of Advanced Medical Materials and Devices, ShanghaiTech University, Shanghai 201210, China · College of Computing and Data Science, Nanyang Technological University, Singapore · School of Automation Science and Electrical Engineering, Beihang University, Beijing 100191, China · State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing 100191, China · Shanghai United Imaging Intelligence, Shanghai 200230, China · Shanghai Clinical Research and Trial Center, Shanghai 201210, China
Brain network analysis provides an interpretable framework for characterizing brain organization and has been widely used for neurological disorder identification. Recent advances in self-supervised learning have motivated the development of brain network foundation models. However, existing approaches are often limited by atlas dependency, insufficient exploitation of multiple network views, and weak incorporation of anatomical priors. In this work, we propose MV-BrainFM, a multi-view brain network foundation model designed to learn generalizable and scalable representations from brain networks constructed with arbitrary atlases. MV-BrainFM explicitly incorporates anatomical distance information into Transformer-based modeling to guide inter-regional interactions, and introduces an unsupervised cross-view consistency learning strategy to align representations from multiple atlases of the same subject in a shared latent space. By jointly enforcing within-view robustness and cross-view alignment during pretraining, the model effectively captures complementary information across heterogeneous network views while remaining atlas-aware. In addition, MV-BrainFM adopts a unified multi-view pretraining paradigm that enables simultaneous learning from multiple datasets and atlases, significantly improving computational efficiency compared to conventional sequential training strategies. The proposed framework also demonstrates strong scalability, consistently benefiting from increasing data diversity while maintaining stable performance across unseen atlas configurations. Extensive experiments on more than 20K subjects from 17 fMRI datasets show that MV-BrainFM consistently outperforms 14 existing brain network foundation models and task-specific baselines under both single-atlas and multi-atlas settings.
Figures & tables
Fig. 1: Comparison of representative brain foundation models.
Fig. 2: The architecture of our proposed MV-BrainFM using Schaefer100 and AAL116 as example. (a) Multiple brain network datasets with various atlases are utilized for pretraining. (b) Datasets with arbitrary atlases can be introduced to finetune the model for downstream tasks with a simple prediction head. (c) The proposed Distance-Biased Transformer (DBT) Encoder.
Datasets
Sex (F/M)
Age (mean ± std)
Class#
Subjects#
Atlases
Pretrain
HBN [ 2 ]
NA
NA
-
12427
Schaefer200, AAL116, Craddock200
ADHD [ 12 ]
383/230
12.1 ± 3.3
-
613
Schaefer100, Schaefer200, Schaefer500
ADNI [ 16 ]
610/552
71.8 ± 6.9
-
1162
Schaefer100, Schaefer200, Schaefer500, AAL116
ABIDE [ 14 ]
122/697
16.9 ± 7.9
-
819
Schaefer100, Schaefer200, Schaefer500, AAL116
PPMI [ 43 ]
82/127
62.9 ± 9.5
-
209
Schaefer100, Schaefer200, Schaefer500, AAL116
Taowu [ 43 ]
17/23
65.0 ± 5.0
-
40
Schaefer100, Schaefer200, Schaefer500, AAL116
TABLE I: Statistics and class information of brain network datasets used in this work. Renji, SMHC, DAMOMRI, and Huashan are private datasets, while the rest are publicly available.
Method
Pretrained
Internal Tests
External Tests
ADHD
ADNI
ABIDE
Avg.
HCP-Gender
HCP-Age
Huashan
Avg.
GCN [ 32 ]
✗
54.66 ± 6.38
66.06 ± 2.20
62.58 ± 4.95
61.1
74.13 ± 4.99
54.43 ± 8.25
54.73 ± 3.20
61.1
BrainNetCNN [ 30 ]
✗
55.10 ± 5.32
68.91 ± 2.06
67.18 ± 7.67
63.73
73.35 ± 14.08
54.81 ± 8.24
59.20 ± 4.73
62.46
BrainGNN [ 35 ]
✗
55.86 ± 5.74
63.23 ± 3.10
60.41 ± 4.02
59.83
72.17 ± 10.30
52.98 ± 5.41
55.82 ± 4.85
60.32
BNT [ 29 ]
✗
53.08 ± 7.21
61.04 ± 6.78
64.46 ± 7.22
59.53
76.57 ± 11.21
51.26 ± 5.98
59.91 ± 3.68
62.58
ContrastPool [ 63 ]
✗
59.65 ± 5.19
64.65 ± 3.66
68.69 ± 2.72
64.33
80.00 ± 5.80
56.30 ± 5.58
58.35 ± 6.11
64.88
TABLE II: Results (average AUC ± standard deviation) on 6 datasets by averaging all available atlases. The best result is highlighted in bold while the runner-up is highlighted in underlined .
Fig. 3: Results on all 4 atlases. Note that BASC122 is the external atlas, which the models have never seen during pretraining.
Fig. 4: Scaling behavior with increasing pretraining atlases. We progressively increase the number of atlases used during pretraining and evaluate the downstream performance on 6 tasks using the Schaefer100 atlas. The results show that MV-BrainFM consistently benefits from larger and more diverse pretraining data, demonstrating clearer scaling behavior than BrainGFM.
Method
Internal Tests
External Tests
ABIDE
ADNI
HCP-Gender
HCP-Age
MGRL [ 13 ]
66.99 ± 4.86
59.57 ± 1.75
76.01 ± 8.40
56.78 ± 6.17
METAFormer [ 41 ]
75.87 ± 3.39
66.40 ± 1.24
84.17 ± 4.39
61.40 ± 4.69
AIDFusion [ 66 ]
75.19 ± 1.38
66.59 ± 2.43
85.34 ± 5.78
59.77 ± 5.35
MV-BrainFM
77.93 ± 2.36
72.63 ± 0.90
86.64 ± 4.54
61.47 ± 5.47
TABLE III: Multi-atlas AUC performance comparison using four atlases (Schaefer100, Schaefer200, AAL116, and BASC122). MV-BrainFM achieves superior performance over existing multi-atlas baselines on all datasets.
Fig. 5: AUC performance of MV-BrainFM using the Schaefer100 atlas under different hidden dimensions (0.95M–31.35M parameters). Increasing model capacity improves performance from very small models, but larger models do not consistently yield additional gains, indicating that data scale may be a more critical factor than parameter scale.
Fig. 6: Visualization of the top 20 edges with the largest distance bias for each attention head. Different heads consistently focus on connections at distinct spatial scales.
Fig. 7: Visualization of salient brain regions identified by MV-BrainFM across different neurological disorders. The highlighted regions correspond to ROIs with the highest attention weights and represent potential disease-related biomarkers for ADHD, ASD (ABIDE), AD/MCI (ADNI and Huashan).
Models
#Params
Pretrain
Finetune Time (s)
#Epoch
Time (s)
ADHD
ABIDE
ADNI
HCP-Gender
HCP-Age
Huashan
BrainLM
12.78M
50
3,989
430
609
785
632
605
462
Brain-JEPA
21.69M
50
4,124
1,038
1,573
1,814
1,630
1,275
1,126
PTGB
0.55M
70
1,574
23
15
22
26
30
29
BrainMass
29.35M
100
6,988
74
80
102
102
117
84
BrainGFM
10.31M
20
3,709
46
58
83
63
63
47
TABLE IV: Time cost (seconds) and number of parameters for brain network foundation models on datasets with Schaefer100 atlas. The finetune time reports the 50 epochs running time for a single fold including training, validation, and test.
Atlas Encoder
Distance Bias
Lrec
Lcc
Lent
ADHD
ABIDE
ADNI
HCP-Gender
HCP-Age
Huashan
✓
✓
54.70 ± 7.76
69.72 ± 1.06
64.23 ± 1.57
83.52 ± 1.79
58.70 ± 6.01
60.06 ± 4.40
✓
✓
✓
57.26 ± 6.47
73.38 ± 1.68
67.74 ± 1.34
85.54 ± 2.52
57.99 ± 5.88
60.71 ± 6.23
✓
✓
✓
✓
58.10 ± 4.97
74.38 ± 2.49
68.20 ± 2.10
85.36 ± 2.67
58.72 ± 6.22
61.61 ± 4.48
✓
✓
✓
✓
58.99 ± 6.45
75.27 ± 2.49
69.24 ± 1.63
86.77 ± 1.77
59.67 ± 5.42
61.31 ± 5.08
✓
✓
✓
✓
61.18 ± 3.54
72.79 ± 1.48
66.19 ± 2.79
84.41 ± 2.05
59.11 ± 4.21
61.64 ± 6.43
✓
✓
✓
✓
59.76 ± 1.85
73.22 ± 1.96
69.30 ± 0.89
86.38 ± 2.90
59.01 ± 5.15
61.83 ± 5.22
TABLE V: Ablation study on components in MV-BrainFM on datasets with the Schaefer100 atlas.
Fig. 8: Hyperparameter sensitivity analysis of MV-BrainFM with respect to the number of attention heads and prototypes across multiple datasets.
Functional magnetic resonance imaging (fMRI) foundation models learn reusable brain representations for prediction, transfer, and neuroscience discovery.
Functional connectivity (FC) derived from resting-state fMRI is widely used to characterize large-scale brain network alterations in neurological and psychiatric disorders. However, FC construction critically depends on the choice of brain atlas, and different parcellations may emphasize distinct organizational features, leading to heterogeneous and sometimes inconsistent representations. Existing multi-atlas approaches partially alleviate this issue but often fuse atlas-derived features or predictions at a relatively shallow level, while single-atlas disentanglement methods do not explicitly address cross-atlas heterogeneity. We propose Multi-Atlas Disentangled Connectivity LEarning (MADCLE), a multi-branch representation learning framework that jointly encodes FC matrices derived from different brain atlases. Rather than introducing a single explicitly shared latent variable across parcellations, MADCLE learns atlas-wise disease-related representations and encourages them to be cross-atlas consistent through distributional alignment. Meanwhile, covariate-related and atlas-dependent residual factors are modeled separately using covariate similarity supervision, atlas-specific reconstruction, and decorrelation constraints, thereby reducing the leakage of non-disease and parcellation-dependent information into the disease-related embeddings. Experiments on the ADNI and ADHD-200 datasets suggest that MADCLE achieves competitive or improved performance compared with single-atlas baselines, multi-atlas GNN/Transformer models, and recent multi-atlas consistency frameworks. These results support the potential value of structured disentanglement for FC-based disorder identification under heterogeneous parcellation schemes.
Minheng Chen, Chao Cao, Jing Zhang +2
Department of Computer Science and Engineering, University of Texas at Arlington, United States · School of Computing, University of Georgia, United States
Foundation models pretrained using self-supervised learning have transformed computer vision by learning transferable representations from large-scale unlabeled data. However, existing foundation models for neuroimaging remain limited by task-specific training, slice-based learning strategies, or relatively small pretraining datasets, restricting their generalizability across diverse brain MRI applications. In this work, we present BrainNext, a general-purpose self-supervised foundation model for volumetric brain MRI analysis. BrainNext combines masked autoencoder (MAE) pretraining with a native three-dimensional Bi-Directional xLSTM-UNet architecture to learn rich anatomical representations from 60,551 unlabeled brain MRI examinations spanning multiple MRI modalities. The pretrained model is subsequently adapted to downstream tasks through lightweight task-specific fine-tuning. We evaluate BrainNext on the Foundation Models for Medical Imaging (FOMO) 2025 Method Track, encompassing classification, segmentation, and brain-age estimation, where it achieved second place overall and ranked first in the meningioma segmentation task on the official FOMO 2025 challenge leaderboard, demonstrating strong transferability across heterogeneous neuroimaging tasks. These results highlight the potential of large-scale self-supervised pretraining to learn robust and transferable volumetric representations, establishing BrainNext as a scalable foundation model for diverse brain MRI applications.
Moona Mazher, Abdul Qayyum, Steven A. Niederer +1
UCL Hawkes Institute, Department of Computer Science, University College London, London, United Kingdom · National Heart & Lung Institute, Imperial College London, London, United Kingdom · Division of Cardiovascular Medicine, Stanford Cardiovascular Institute, Stanford University, California, USA
We study the problem of training self-supervised foundation models for functional MRI. Our main contributions are: (1) we introduce a new model family (CortexMAE) trained using the masked autoencoder framework on 2.1K hours of open fMRI data, and (2) we release the first open evaluation suite (Brainmarks) for fMRI foundation models. Our core innovation is simple: we adapt the Vision Transformer to fMRI by first converting each 3D fMRI volume to a 2D map using a cortical flat map projection. We directly compare flat maps to both parcellation and volume-based representations. While each has its advantages, flat maps generally perform best. We perform the first systematic scaling analysis for fMRI and observe strict power law scaling, albeit with limits. Finally, we use Brainmarks to do controlled benchmark comparisons. On subject-level trait prediction, we report a challenging null result: no single model achieves clear state-of-the-art performance. Moreover, all models struggle to outperform a simple functional connectivity baseline. On cognitive state decoding, we observe more robust performance, and in this setting our CortexMAE family outperforms prior models by a large margin. Code, models, and datasets are available at https://github.com/MedARC-AI/CortexMAE and https://github.com/MedARC-AI/Brainmarks.