- Rating
- 2
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper proposes CoFlu-Brain, a self-supervised foundation model for resting-state functional connectivity. They extend the standard static FC matrix to a Fisher-z FC matrix, with the average of framewise cofluctuation products. The model is pretrained on 13444 sbujects from eight cohorts and evaluation on nine tasks in seven datasets.
Critical strengths
-
The core idea is simple: treat an FC matrix as a finite-sample estimate and building views from complementary frame subsets.
-
The age prediction improves substantially across the datasets.
Critical weaknesses
-
The method recombines existing works. Each “cofluctuation-based FC view” is simply Pearson FC computed on a random frame subset. That is the timepoint-dropping augmentation BrainMass already uses. The consistency term is LeJEPA’s view-to-center prediction loss and SIGReg is taken from LeJEPA unchanged, so the only new component is a predictor from each retained view to its complementary dropped view.
-
The framing of cofluctuation is problematic. Averaging framewise products over a frame subset reproduces Pearson correlation exactly, so the edge-time-series view adds notation but no new signal.
-
There is no basic baseline like using the raw Schaefer-400 FC vector under the same linear probe, or randomly intialized encoder.
-
The evaluation is problematic. CoFlu-Brain reads 400 ROIs while the baselines read 100, 116 or 424 regions through their own feature paths and trained on different corpus.
-
The performance gains are marginal. On five of the nine tasks, CoFlu-Brain’s margin over the best baseline is under 0.01 weighted F1 (for example, ADNI AD .6835 vs .6832).
-
The paper claims that the method is a foundation model. But the evaluation is limited to frozen linear probes on static FC, with no fine-tuning, few-shot, or cross-atlas evaluation. The other foundation models that take in different inputs, like Brain-JEPA and Omni-fMRI are not compared. It's also questionable to work in the FC while current fMRI foundation models shift to atlas-free.
Questions for authors
-
How does a linear probe on the vectorized Schaefer-400 Fisher-z FC of the same 200-frame segments compare, using the same folds and probes? How does a randomly initialized CoFlu-Brain encoder compare?
-
What is the significance of the reported comparisons?
-
Do the results change when downstream FC uses all available frames instead of one 200-frame segment, for CoFlu-Brain and the baselines alike?
-
Why does the ABIDE evaluation include only 173 participants, and how were the ADNI subsets selected?
-
How correlated are the retained and dropped FC matrices? Does the advantage of random partitioning hold at each cohort’s native TR, or with random contiguous chunks longer than the hemodynamic response?
11 comments
Pretraining versus random initialization.
We compared pretrained and randomly initialized encoders across three seeds using identical architectures, inputs, participant folds, and fixed linear probes. All encoders were frozen and evaluated on the full downstream pool using participant-grouped five-fold cross-validation. Pretraining yields better mean performance on 14 of 15 tasks, with HCP sex weighted F1 of 0.7746 versus 0.7352 (Reply Table 3).
Table 3: Pretrained versus randomly initialized encoders
Means are averaged over three seeds and five participant-grouped folds; ± denotes fold SD after averaging seeds. Both conditions use identical inputs and folds from the full downstream pool.
| Task / metric | Pretrained | Random init |
|---|---|---|
| HCP / Sex / weighted F1 ↑ | 0.7746 ± 0.0105 | 0.7352 ± 0.0149 |
| ABCD / Sex / weighted F1 ↑ | 0.5961 ± 0.0200 | 0.5474 ± 0.0157 |
| NKI / Age / nMAE ↓ | 0.5628 ± 0.0274 | 0.7483 ± 0.0125 |
| NKI / Education / weighted F1 ↑ | 0.6439 ± 0.0180 | 0.5327 ± 0.0143 |
| NKI / Sex / weighted F1 ↑ | 0.6583 ± 0.0300 | 0.6121 ± 0.0142 |
| NKI / BMI / nMAE ↓ | 0.7566 ± 0.0431 | 0.8573 ± 0.0300 |
| ABIDE / Age / nMAE ↓ | 1.0317 ± 0.1298 | 1.4893 ± 0.1604 |
| ABIDE / ASD / weighted F1 ↑ | 0.5567 ± 0.0476 | 0.5515 ± 0.0697 |
| ADHD200 / Age / nMAE ↓ | 0.6557 ± 0.0370 | 0.7733 ± 0.0253 |
| ADHD200 / Sex / weighted F1 ↑ | 0.5999 ± 0.0290 | 0.6038 ± 0.0113 |
| ADHD200 / Diagnosis / weighted F1 ↑ | 0.5889 ± 0.0150 | 0.5380 ± 0.0181 |
| SALD / Age / nMAE ↓ | 0.6571 ± 0.0431 | 0.8960 ± 0.0460 |
| PPMI / Group / weighted F1 ↑ | 0.5775 ± 0.0166 | 0.5311 ± 0.0382 |
| ADNI / AD / weighted F1 ↑ | 0.6835 ± 0.0307 | 0.6570 ± 0.0608 |
| ADNI / MCI / weighted F1 ↑ | 0.5538 ± 0.0441 | 0.5037 ± 0.0557 |
3. Parcellations and comparison settings (W4):
Thank you for raising the question of parcellation differences across methods.
Baselines with at least 400 regions.
We added Brain-JEPA (450 regions: Schaefer-400 plus Tian-50) and Brain-DiT (424 regions) to manuscript Table 1, extending the comparison to baselines with at least 400 regions under a common participant-split and linear-probe protocol (Reply Table 4).
Table 4: Added 400+ region baselines (manuscript Table 1)
| Task / metric | CoFlu-Brain | Brain-JEPA (450) | Brain-DiT (424) | FC mean |
|---|---|---|---|---|
| HCP / Sex / weighted F1 ↑ | 0.7745 ± 0.0103 | 0.5766 ± 0.0207 | 0.7187 ± 0.0135 | 0.7304 ± 0.0243 |
| ABCD / Sex / weighted F1 ↑ | 0.5961 ± 0.0190 | 0.5189 ± 0.0202 | 0.5575 ± 0.0330 | 0.5797 ± 0.0295 |
| PPMI / Group / weighted F1 ↑ | 0.5775 ± 0.0173 | 0.5327 ± 0.0294 | 0.5194 ± 0.0330 | 0.5455 ± 0.0458 |
| ADNI / AD / weighted F1 ↑ | 0.6835 ± 0.0307 | 0.5589 ± 0.0573 | 0.6251 ± 0.0648 | 0.6137 ± 0.0255 |
| ADNI / MCI / weighted F1 ↑ | 0.5538 ± 0.0441 | 0.5010 ± 0.0475 | 0.5160 ± 0.0480 | 0.5238 ± 0.0309 |
| ABIDE / ASD / weighted F1 ↑ | 0.5567 ± 0.0476 | 0.5048 ± 0.0461 | 0.5884 ± 0.0504 | 0.5569 ± 0.0830 |
| ADHD200 / Diagnosis / weighted F1 ↑ | 0.5891 ± 0.0151 | 0.5231 ± 0.0286 | 0.5867 ± 0.0220 | 0.5397 ± 0.0210 |
| SALD / Age / nMAE ↓ | 0.6571 ± 0.0431 | 1.1583 ± 0.1500 | 0.8813 ± 0.0432 | 1.2716 ± 0.1386 |
| ADHD200 / Age / nMAE ↓ | 0.6557 ± 0.0370 | 0.9341 ± 0.0600 | 0.7794 ± 0.0438 | 0.7667 ± 0.0486 |
| Schaefer-100 transfer. |
We also evaluated Schaefer-100, -200, and -400. With both methods using Schaefer-100 and the same downstream participant splits and linear-probe protocol, CoFlu-Brain achieves better primary-metric means than BrainMASS on seven of nine tasks, showing that its transfer advantage is not solely attributable to a finer parcellation (Reply Table 5).
Table 5: CoFlu-Brain versus BrainMASS with Schaefer-100 (five-fold mean ± SD)
The best mean in each row is bolded.
| Task / metric | BrainMASS (S100) | CoFlu-Brain (S100) |
|---|---|---|
| HCP / Sex / weighted F1 ↑ | 0.7203 ± 0.0048 | 0.7551 ± 0.0158 |
| ABCD / Sex / weighted F1 ↑ | 0.6215 ± 0.0515 | 0.6041 ± 0.0256 |
| PPMI / Group / weighted F1 ↑ | 0.5751 ± 0.0301 | 0.5811 ± 0.0265 |
| ADNI / AD / weighted F1 ↑ | 0.5811 ± 0.0393 | 0.6783 ± 0.0775 |
| ADNI / MCI / weighted F1 ↑ | 0.5045 ± 0.0346 | 0.5498 ± 0.0660 |
| ABIDE / ASD / weighted F1 ↑ | 0.5472 ± 0.0527 | 0.5300 ± 0.0851 |
| ADHD200 / Diagnosis / weighted F1 ↑ | 0.5549 ± 0.0254 | 0.5711 ± 0.0182 |
| ADHD200 / Age / nMAE ↓ | 0.9328 ± 0.0377 | 0.6759 ± 0.0177 |
| SALD / Age / nMAE ↓ | 0.9064 ± 0.0877 | 0.7744 ± 0.0462 |
Compact readout.
Its mean-pooled readout is 512-dimensional: one quarter of LCM’s 2,048 dimensions, and smaller than BrainLM’s 1,536 and BrainMASS’s 800. These results demonstrate competitive transfer with both a coarser parcellation and a compact representation (Reply Table 6).
Table 6: Readout dimensions
| Model | Input regions | Readout dimensions |
|---|---|---|
| CoFlu-Brain S100 | 100 | 512 |
| LCM | 116 | 2048 |
| BrainLM | 424 | 1536 |
| BrainMASS | 100 | 800 |
Choice of resolution.
Finer parcellations provide finer spatial information. Schaefer-400 has better means than Schaefer-200 on six of nine tasks and the best mean among the three resolutions on five, supporting Schaefer-400 as our default resolution (Reply Table 7).
Table 7: Parcellation resolution (five-fold mean ± SD)
The best mean in each row is bolded.
| Task / metric | S100 | S200 | S400 |
|---|---|---|---|
| HCP / Sex / weighted F1 ↑ | 0.7551 ± 0.0158 | 0.7561 ± 0.0256 | 0.7875 ± 0.0115 |
| ABCD / Sex / weighted F1 ↑ | 0.6041 ± 0.0256 | 0.5793 ± 0.0333 | 0.5889 ± 0.0178 |
| PPMI / Group / weighted F1 ↑ | 0.5811 ± 0.0265 | 0.5844 ± 0.0170 | 0.5802 ± 0.0206 |
| ADNI / AD / weighted F1 ↑ | 0.6783 ± 0.0775 | 0.7246 ± 0.1044 | 0.7219 ± 0.0657 |
| ADNI / MCI / weighted F1 ↑ | 0.5498 ± 0.0660 | 0.5718 ± 0.0556 | 0.5461 ± 0.0553 |
| ABIDE / ASD / weighted F1 ↑ | 0.5300 ± 0.0851 | 0.5856 ± 0.0424 | 0.6013 ± 0.1107 |
| ADHD200 / Diagnosis / weighted F1 ↑ | 0.5711 ± 0.0182 | 0.5826 ± 0.0216 | 0.5872 ± 0.0215 |
| ADHD200 / Age / nMAE ↓ | 0.6759 ± 0.0177 | 0.6683 ± 0.0407 | 0.6523 ± 0.0315 |
| SALD / Age / nMAE ↓ | 0.7744 ± 0.0462 | 0.6999 ± 0.0673 | 0.6981 ± 0.0586 |
4. Magnitude and significance (W5, Q2):
We appreciate your request for a clearer assessment of the overall gains and statistical evidence.
CoFlu-Brain achieves the best cohort-averaged rank (1.43) among eight methods. The overall rank permutation test detects differences among methods (p = 1.15 × 10⁻⁴). Together, these results support CoFlu-Brain’s overall competitiveness across downstream tasks (Reply Table 8).
Table 8: Cohort-averaged ranks
| Model | Cohort-averaged rank ↓ |
|---|---|
| CoFlu-Brain | 1.43 |
| FC mean readout | 3.93 |
| BrainMASS | 4.00 |
| Brain-DiT | 4.14 |
| LCM | 4.43 |
| BrainGFM | 4.64 |
| BrainLM | 6.36 |
| Brain-JEPA | 7.07 |
5. Evaluation scope and FC input (W6):
Thank you for the questions about the evaluation scope and the choice of FC input.
Full fine-tuning is sensitive to learning rate, regularization, and training duration; few-shot results also depend on annotation budget and sample selection. These evaluations combine pretraining and downstream adaptation effects, making it difficult to isolate pretrained-representation quality. We therefore use frozen-encoder linear probing to more directly assess cross-cohort representation transfer under common participant splits and probe settings. Added Brain-JEPA/Brain-DiT comparisons and Schaefer-100/200/400 experiments broaden the input and resolution coverage.
FC explicitly represents interregional functional coupling in a compact matrix, making it suitable for large-scale multi-cohort pretraining and connectivity-representation transfer. Compared with voxel-level spatiotemporal modeling, regional representations reduce input size and computational demands while retaining coupling information directly relevant to brain-network analysis. We therefore use FC as the modeling unit and investigate self-supervised learning from connectivity estimates formed by different frame compositions. FC and atlas-free models offer different spatial granularities and computational costs.
6. Temporal coverage (Q3):
Thank you for suggesting that we examine the effect of temporal coverage.
We compared one-window representations with eight-window feature averaging on six tasks, using the same scans, model, folds, and probes. Means improve on five tasks: HCP sex F1, 0.7644 to 0.7978; PPMI group F1, 0.5457 to 0.5903; SALD age nMAE, 0.6981 to 0.6175. In the main comparison, all models use the same 200-frame source segments as their input basis to maintain consistent downstream temporal coverage (Reply Table 9).
Table 9: Temporal-coverage comparison (five-fold mean ± SD)
| Task / metric | One window | Eight-window mean |
|---|---|---|
| HCP / Sex / weighted F1 ↑ | 0.7644 ± 0.0175 | 0.7978 ± 0.0375 |
| SALD / Age / nMAE ↓ | 0.6981 ± 0.0586 | 0.6175 ± 0.0724 |
| PPMI / Group / weighted F1 ↑ | 0.5457 ± 0.0246 | 0.5903 ± 0.0547 |
| ADNI / AD / weighted F1 ↑ | 0.7219 ± 0.0657 | 0.7257 ± 0.0829 |
| ADNI / MCI / weighted F1 ↑ | 0.5461 ± 0.0553 | 0.5465 ± 0.0470 |
| ABIDE / ASD / weighted F1 ↑ | 0.6013 ± 0.1107 | 0.5868 ± 0.1151 |
Thank you for the detailed feedback. We clarify the design and summarize the added evaluations below.
1. Contribution and cofluctuation framing (W1, W2):
We appreciate your comments on the task design and its connection to the cofluctuation perspective.
BrainMASS drops time points to construct FC augmentations and learns through masked modeling and representation alignment. CoFlu-Brain additionally turns the dropped frames into FC prediction targets, combining retained-to-dropped representation prediction with consistency across retained views. Frame partitioning therefore both generates augmented views and defines a latent prediction task between retained and dropped observations. The cofluctuation perspective explains the relationship between frame partitioning and FC construction; we revised the wording to make the task-design contribution precise. Appendix D compares random complementary, contiguous, configuration-similar, configuration-diverse, and independent sampling under matched training settings. Random complementary sampling ranks in the top two on seven of nine tasks and has better means than independent sampling on six. These results provide further empirical support for the proposed view construction in downstream transfer (Reply Table 1).
Table 1: View-construction comparison (five-fold mean ± SD)
The two best means in each row are bolded.
2. Basic baselines (W3, Q1):
Thank you for recommending these controls to assess the benefit of pretraining.
Direct FC readout.
Manuscript Table 1 now includes an FC mean-readout baseline requiring no pretraining: 400-dimensional features from the same Schaefer-400 Fisher-z FC matrices and 200-frame segments, evaluated with identical participant splits and linear probes. CoFlu-Brain has better means on eight of nine tasks: HCP sex F1, 0.7745 versus 0.7304; ADNI AD F1, 0.6835 versus 0.6137; ADHD age nMAE, 0.6557 versus 0.7667 (Reply Table 2).
Table 2: FC mean-readout comparison (manuscript Table 1)