acceptodds
Under review as a conference paper at ICLR 2027

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

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.

Primary area:
applications to neuroscience & cognitive science
Keywords:
fMRI, functional connectivity, framewise cofluctuation, self-supervised learning, brain foundation model

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_WBhoOpenReview ↗
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
  1. The core idea is simple: treat an FC matrix as a finite-sample estimate and building views from complementary frame subsets.

  2. The age prediction improves substantially across the datasets.

Critical weaknesses
  1. 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.

  2. 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.

  3. There is no basic baseline like using the raw Schaefer-400 FC vector under the same linear probe, or randomly intialized encoder.

  4. 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.

  5. 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).

  6. 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
  1. 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?

  2. What is the significance of the reported comparisons?

  3. Do the results change when downstream FC uses all available frames instead of one 200-frame segment, for CoFlu-Brain and the baselines alike?

  4. Why does the ABIDE evaluation include only 173 participants, and how were the ADNI subsets selected?

  5. 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

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.

Task / metricRandom complementContiguousSimilarDiverseIndependent
HCP / Sex / weighted F1 ↑0.7744 ± 0.01540.7655 ± 0.01020.7708 ± 0.02360.7730 ± 0.01470.7406 ± 0.0081
ABCD / Sex / weighted F1 ↑0.6055 ± 0.02980.5478 ± 0.04480.5850 ± 0.01650.6077 ± 0.03110.5809 ± 0.0269
PPMI / Group / weighted F1 ↑0.5679 ± 0.05070.5323 ± 0.03660.5815 ± 0.03470.5867 ± 0.04130.5598 ± 0.0265
ADNI / AD / weighted F1 ↑0.6506 ± 0.07800.6802 ± 0.10720.6605 ± 0.07510.6701 ± 0.04210.6784 ± 0.0561
ADNI / MCI / weighted F1 ↑0.5812 ± 0.03940.5798 ± 0.05440.5471 ± 0.07700.5538 ± 0.02400.5916 ± 0.0342
ABIDE / ASD / weighted F1 ↑0.5743 ± 0.03690.5551 ± 0.04500.5184 ± 0.11160.5573 ± 0.07260.6073 ± 0.0495
ADHD200 / Diagnosis / weighted F1 ↑0.5886 ± 0.03560.5774 ± 0.03070.5779 ± 0.02390.5842 ± 0.01120.5602 ± 0.0305
ADHD200 / Age / nMAE ↓0.6663 ± 0.05260.6820 ± 0.03500.6323 ± 0.02330.6714 ± 0.04970.6688 ± 0.0321
SALD / Age / nMAE ↓0.6974 ± 0.02380.7905 ± 0.04360.7340 ± 0.01810.7836 ± 0.06370.7225 ± 0.0219

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)

Task / metricCoFlu-BrainFC mean
HCP / Sex / weighted F1 ↑0.7745 ± 0.01030.7304 ± 0.0243
ABCD / Sex / weighted F1 ↑0.5961 ± 0.01900.5797 ± 0.0295
PPMI / Group / weighted F1 ↑0.5775 ± 0.01730.5455 ± 0.0458
ADNI / AD / weighted F1 ↑0.6835 ± 0.03070.6137 ± 0.0255
ADNI / MCI / weighted F1 ↑0.5538 ± 0.04410.5238 ± 0.0309
ABIDE / ASD / weighted F1 ↑0.5567 ± 0.04760.5569 ± 0.0830
ADHD200 / Diagnosis / weighted F1 ↑0.5891 ± 0.01510.5397 ± 0.0210
SALD / Age / nMAE ↓0.6571 ± 0.04311.2716 ± 0.1386
ADHD200 / Age / nMAE ↓0.6557 ± 0.03700.7667 ± 0.0486

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 / metricPretrainedRandom init
HCP / Sex / weighted F1 ↑0.7746 ± 0.01050.7352 ± 0.0149
ABCD / Sex / weighted F1 ↑0.5961 ± 0.02000.5474 ± 0.0157
NKI / Age / nMAE ↓0.5628 ± 0.02740.7483 ± 0.0125
NKI / Education / weighted F1 ↑0.6439 ± 0.01800.5327 ± 0.0143
NKI / Sex / weighted F1 ↑0.6583 ± 0.03000.6121 ± 0.0142
NKI / BMI / nMAE ↓0.7566 ± 0.04310.8573 ± 0.0300
ABIDE / Age / nMAE ↓1.0317 ± 0.12981.4893 ± 0.1604
ABIDE / ASD / weighted F1 ↑0.5567 ± 0.04760.5515 ± 0.0697
ADHD200 / Age / nMAE ↓0.6557 ± 0.03700.7733 ± 0.0253
ADHD200 / Sex / weighted F1 ↑0.5999 ± 0.02900.6038 ± 0.0113
ADHD200 / Diagnosis / weighted F1 ↑0.5889 ± 0.01500.5380 ± 0.0181
SALD / Age / nMAE ↓0.6571 ± 0.04310.8960 ± 0.0460
PPMI / Group / weighted F1 ↑0.5775 ± 0.01660.5311 ± 0.0382
ADNI / AD / weighted F1 ↑0.6835 ± 0.03070.6570 ± 0.0608
ADNI / MCI / weighted F1 ↑0.5538 ± 0.04410.5037 ± 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 / metricCoFlu-BrainBrain-JEPA (450)Brain-DiT (424)FC mean
HCP / Sex / weighted F1 ↑0.7745 ± 0.01030.5766 ± 0.02070.7187 ± 0.01350.7304 ± 0.0243
ABCD / Sex / weighted F1 ↑0.5961 ± 0.01900.5189 ± 0.02020.5575 ± 0.03300.5797 ± 0.0295
PPMI / Group / weighted F1 ↑0.5775 ± 0.01730.5327 ± 0.02940.5194 ± 0.03300.5455 ± 0.0458
ADNI / AD / weighted F1 ↑0.6835 ± 0.03070.5589 ± 0.05730.6251 ± 0.06480.6137 ± 0.0255
ADNI / MCI / weighted F1 ↑0.5538 ± 0.04410.5010 ± 0.04750.5160 ± 0.04800.5238 ± 0.0309
ABIDE / ASD / weighted F1 ↑0.5567 ± 0.04760.5048 ± 0.04610.5884 ± 0.05040.5569 ± 0.0830
ADHD200 / Diagnosis / weighted F1 ↑0.5891 ± 0.01510.5231 ± 0.02860.5867 ± 0.02200.5397 ± 0.0210
SALD / Age / nMAE ↓0.6571 ± 0.04311.1583 ± 0.15000.8813 ± 0.04321.2716 ± 0.1386
ADHD200 / Age / nMAE ↓0.6557 ± 0.03700.9341 ± 0.06000.7794 ± 0.04380.7667 ± 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 / metricBrainMASS (S100)CoFlu-Brain (S100)
HCP / Sex / weighted F1 ↑0.7203 ± 0.00480.7551 ± 0.0158
ABCD / Sex / weighted F1 ↑0.6215 ± 0.05150.6041 ± 0.0256
PPMI / Group / weighted F1 ↑0.5751 ± 0.03010.5811 ± 0.0265
ADNI / AD / weighted F1 ↑0.5811 ± 0.03930.6783 ± 0.0775
ADNI / MCI / weighted F1 ↑0.5045 ± 0.03460.5498 ± 0.0660
ABIDE / ASD / weighted F1 ↑0.5472 ± 0.05270.5300 ± 0.0851
ADHD200 / Diagnosis / weighted F1 ↑0.5549 ± 0.02540.5711 ± 0.0182
ADHD200 / Age / nMAE ↓0.9328 ± 0.03770.6759 ± 0.0177
SALD / Age / nMAE ↓0.9064 ± 0.08770.7744 ± 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

ModelInput regionsReadout dimensions
CoFlu-Brain S100100512
LCM1162048
BrainLM4241536
BrainMASS100800

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 / metricS100S200S400
HCP / Sex / weighted F1 ↑0.7551 ± 0.01580.7561 ± 0.02560.7875 ± 0.0115
ABCD / Sex / weighted F1 ↑0.6041 ± 0.02560.5793 ± 0.03330.5889 ± 0.0178
PPMI / Group / weighted F1 ↑0.5811 ± 0.02650.5844 ± 0.01700.5802 ± 0.0206
ADNI / AD / weighted F1 ↑0.6783 ± 0.07750.7246 ± 0.10440.7219 ± 0.0657
ADNI / MCI / weighted F1 ↑0.5498 ± 0.06600.5718 ± 0.05560.5461 ± 0.0553
ABIDE / ASD / weighted F1 ↑0.5300 ± 0.08510.5856 ± 0.04240.6013 ± 0.1107
ADHD200 / Diagnosis / weighted F1 ↑0.5711 ± 0.01820.5826 ± 0.02160.5872 ± 0.0215
ADHD200 / Age / nMAE ↓0.6759 ± 0.01770.6683 ± 0.04070.6523 ± 0.0315
SALD / Age / nMAE ↓0.7744 ± 0.04620.6999 ± 0.06730.6981 ± 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

ModelCohort-averaged rank ↓
CoFlu-Brain1.43
FC mean readout3.93
BrainMASS4.00
Brain-DiT4.14
LCM4.43
BrainGFM4.64
BrainLM6.36
Brain-JEPA7.07

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 / metricOne windowEight-window mean
HCP / Sex / weighted F1 ↑0.7644 ± 0.01750.7978 ± 0.0375
SALD / Age / nMAE ↓0.6981 ± 0.05860.6175 ± 0.0724
PPMI / Group / weighted F1 ↑0.5457 ± 0.02460.5903 ± 0.0547
ADNI / AD / weighted F1 ↑0.7219 ± 0.06570.7257 ± 0.0829
ADNI / MCI / weighted F1 ↑0.5461 ± 0.05530.5465 ± 0.0470
ABIDE / ASD / weighted F1 ↑0.6013 ± 0.11070.5868 ± 0.1151
Reviewer_zhTMOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CoFlu-Brain, a self-supervised framework for learning functional-connectivity representations from fMRI. Each input segment is partitioned into multiple retained–dropped frame pairs, and a separate FC matrix is computed for each subset. A shared Transformer encoder is trained using consistency across retained views, prediction of the corresponding dropped-view representations, and SIGReg regularization. The model is pretrained on 13,444 participants from eight dataset groups and evaluated through frozen linear probing on seven classification and two age-prediction tasks. Additional experiments examine temporal partitioning, objective components, data coverage, and training configurations.

Critical strengths
  1. Treating FC as a finite-sample estimate and resampling frames to generate training views is simple and well motivated. The reported pretraining cost of approximately 10.54 GPU-hours also suggests a relatively accessible training setup.

  2. The evaluation covers nine tasks across seven cohorts, including five cohorts excluded from pretraining. Participant-grouped folds, multiple pretraining seeds for the main comparisons, and complete appendix tables improve transparency. The age-prediction results are particularly encouraging.

  3. The seed-averaged random-versus-ordered partitioning comparison favors random partitioning across all seven evaluation cohorts, providing useful empirical support for this design choice.

Critical weaknesses
  1. The “cofluctuation-based FC views” in Section 3.2 are Pearson FC matrices computed on frame subsets. The encoder receives aggregated connectivity rather than frame-resolved edge time series. Since BrainMASS already uses frame-dropping FC augmentation and latent alignment, the incremental benefit of complementary retained–dropped pairing needs stronger controlled validation.

  2. The interpretation of dropped views as complementary evidence requires further clarification. BOLD signals are temporally autocorrelated, and interpolation to 0.72s can make disjoint frame subsets depend on the same acquired volumes. The current experiments do not distinguish the benefit of complementary evidence prediction from that of aligning correlated resampled views.

  3. The baseline comparison uses different parcellations, including Schaefer-400 for CoFlu-Brain and Schaefer-100 for BrainMASS and BrainGFM. Probes also use fixed regularization parameters without task-specific selection. Without a tuned raw-FC baseline and matched probing procedures, the source and magnitude of the representation-learning benefit remain unclear.

  4. The objective ablations use only one pretraining seed. Moreover, “No prediction” appears to retain a consistency schedule that reaches zero at epoch 25, leaving only SIGReg for most of training. This complicates interpretation of the prediction ablation. Several clinical improvements are also small, and their reliability is not established through paired uncertainty estimates.

Questions for authors
  1. Please report a tuned linear baseline on raw Schaefer-400 FC and evaluate CoFlu-Brain and the strongest released baselines using the same training-fold-only procedure for selecting C/α. Reporting actual constant-age and majority-class reference scores would also help contextualize the results.

  2. Were the downstream datasets also interpolated to TR 0.72s, and how many acquired volumes underlie each 200-frame segment? Does the advantage of random partitioning remain when retained and dropped sets are constructed from native acquired volumes rather than interpolated frames?

  3. What does complementary retained–dropped pairing add over independently sampled FC views? Please compare these alternatives while matching source and target frame counts, the encoder, predictor, all loss terms, and the training budget, so that only the pairing strategy changes.

  4. Does the “No prediction” variant retain the consistency schedule that becomes zero at epoch 25? If so, please compare it with consistency active throughout training. Multi-seed objective ablations and participant-level paired uncertainty estimates for the small clinical gains would also strengthen the evidence.

9 comments

Thank you for the detailed comments. We address the related points in four groups.

1. Complementary pairing (W1, Q3):

Thank you for asking us to clarify the role of complementary pairing.

The input is Pearson FC computed on frame subsets. CoFlu-Brain learns relationships between these estimates through retained-view consistency and complementary dropped-view prediction. Appendix D compares complementary and independently sampled views, matching source/target frame counts, initialization, data, encoder, predictor, all losses, and training budget. Complementary pairing has better means on six of nine tasks, including HCP sex F1 (0.7744 versus 0.7406) and SALD age nMAE (0.6974 versus 0.7225), supporting its contribution (Reply Table 1).

Table 1: Five view-construction strategies (five-fold mean ± SD)

The two best means in each row are bolded.

Task / metricRandom complementContiguousSimilarDiverseIndependent
HCP / Sex / weighted F1 ↑0.7744 ± 0.01540.7655 ± 0.01020.7708 ± 0.02360.7730 ± 0.01470.7406 ± 0.0081
ABCD / Sex / weighted F1 ↑0.6055 ± 0.02980.5478 ± 0.04480.5850 ± 0.01650.6077 ± 0.03110.5809 ± 0.0269
PPMI / Group / weighted F1 ↑0.5679 ± 0.05070.5323 ± 0.03660.5815 ± 0.03470.5867 ± 0.04130.5598 ± 0.0265
ADNI / AD / weighted F1 ↑0.6506 ± 0.07800.6802 ± 0.10720.6605 ± 0.07510.6701 ± 0.04210.6784 ± 0.0561
ADNI / MCI / weighted F1 ↑0.5812 ± 0.03940.5798 ± 0.05440.5471 ± 0.07700.5538 ± 0.02400.5916 ± 0.0342
ABIDE / ASD / weighted F1 ↑0.5743 ± 0.03690.5551 ± 0.04500.5184 ± 0.11160.5573 ± 0.07260.6073 ± 0.0495
ADHD200 / Diagnosis / weighted F1 ↑0.5886 ± 0.03560.5774 ± 0.03070.5779 ± 0.02390.5842 ± 0.01120.5602 ± 0.0305
ADHD200 / Age / nMAE ↓0.6663 ± 0.05260.6820 ± 0.03500.6323 ± 0.02330.6714 ± 0.04970.6688 ± 0.0321
SALD / Age / nMAE ↓0.6974 ± 0.02380.7905 ± 0.04360.7340 ± 0.01810.7836 ± 0.06370.7225 ± 0.0219

2. Temporal harmonization and complementarity (W2, Q2):

Thank you for the question. Both our pretraining and downstream data have been preprocessed onto a common temporal grid with TR = 0.72 s, so that 200-frame inputs span a consistent physical timescale across cohorts. Given the rebuttal timeline, we will include comparisons at native TR in future work.

Here, “complementary” refers to disjoint processed-frame subsets whose union covers the segment, without assuming statistical independence. The model learns relationships between the FC estimates formed from these subsets.

3. Baselines and probing (W3, Q1):

Thank you for the suggestions on baselines and the probing protocol.

FC mean-readout baseline.

We added an FC mean-readout baseline with comparable feature dimensionality, using the same participant splits and fixed-regularization probes (Reply Table 2).

Table 2: FC mean-readout comparison (manuscript Table 1)

Task / metricCoFlu-BrainFC mean
HCP / Sex / weighted F1 ↑0.7745 ± 0.01030.7304 ± 0.0243
ABCD / Sex / weighted F1 ↑0.5961 ± 0.01900.5797 ± 0.0295
PPMI / Group / weighted F1 ↑0.5775 ± 0.01730.5455 ± 0.0458
ADNI / AD / weighted F1 ↑0.6835 ± 0.03070.6137 ± 0.0255
ADNI / MCI / weighted F1 ↑0.5538 ± 0.04410.5238 ± 0.0309
ABIDE / ASD / weighted F1 ↑0.5567 ± 0.04760.5569 ± 0.0830
ADHD200 / Diagnosis / weighted F1 ↑0.5891 ± 0.01510.5397 ± 0.0210
SALD / Age / nMAE ↓0.6571 ± 0.04311.2716 ± 0.1386
ADHD200 / Age / nMAE ↓0.6557 ± 0.03700.7667 ± 0.0486

Fixed-probe protocol and regularization.

We deliberately use identical fixed-regularization probes to compare frozen representations under a common readout protocol and limit the influence of downstream hyperparameter selection. In a sensitivity check, changing only probe regularization reduced SALD age nMAE from 0.9064 to 0.6588 for BrainMASS and from 0.6378 to 0.5709 for CoFlu-Brain, with pretrained features unchanged. Thus, a higher probe score alone does not establish a stronger pretrained representation. The fixed-probe protocol provides a consistent basis for comparing representations (Reply Table 3).

Table 3: Probe regularization sensitivity (unchanged pretrained features; one CoFlu-Brain pretraining run; descriptive sensitivity check)

Model / SALD ageFixed-probe nMAERegularization sensitivity nMAE
BrainMASS0.90640.6588
CoFlu-Brain0.63780.5709

Label-only references.

We also evaluated majority-class and constant-age references on the same five folds, determining each reference from training-fold labels only and normalizing age nMAE by the training-fold population SD. CoFlu-Brain outperforms these references on all nine primary metrics: for example, HCP sex F1 is 0.7745 versus 0.3735, and SALD age nMAE is 0.6571 versus 0.8876 (CoFlu-Brain versus reference) (Reply Table 4).

Table 4: Majority-class / constant-age references

TaskMetricCoFlu-Brain meanConstant reference
HCP / SexF1 ↑0.77450.3735 ± 0.0444
ABCD / SexF1 ↑0.59610.3483 ± 0.0493
PPMI / GroupF1 ↑0.57750.4999 ± 0.0705
ADNI / ADF1 ↑0.68350.2863 ± 0.0162
ADNI / MCIF1 ↑0.55380.3109 ± 0.0214
ABIDE / ASDF1 ↑0.55670.3561 ± 0.0186
ADHD200 / DiagnosisF1 ↑0.58910.3760 ± 0.0377
SALD / AgenMAE ↓0.65710.8876 ± 0.0467
ADHD200 / AgenMAE ↓0.65570.8139 ± 0.0678

Coarser parcellation.

The added Brain-JEPA and Brain-DiT baselines use 450 regions (Schaefer-400 plus Tian-50) and 424 regions, respectively. We additionally evaluated Schaefer-100, -200, and -400 variants. 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 / metricBrainMASS (S100)CoFlu-Brain (S100)
HCP / Sex / weighted F1 ↑0.7203 ± 0.00480.7551 ± 0.0158
ABCD / Sex / weighted F1 ↑0.6215 ± 0.05150.6041 ± 0.0256
PPMI / Group / weighted F1 ↑0.5751 ± 0.03010.5811 ± 0.0265
ADNI / AD / weighted F1 ↑0.5811 ± 0.03930.6783 ± 0.0775
ADNI / MCI / weighted F1 ↑0.5045 ± 0.03460.5498 ± 0.0660
ABIDE / ASD / weighted F1 ↑0.5472 ± 0.05270.5300 ± 0.0851
ADHD200 / Diagnosis / weighted F1 ↑0.5549 ± 0.02540.5711 ± 0.0182
ADHD200 / Age / nMAE ↓0.9328 ± 0.03770.6759 ± 0.0177
SALD / Age / nMAE ↓0.9064 ± 0.08770.7744 ± 0.0462

Parcellation resolution.

With 512-dimensional frozen features and the same downstream folds and probes, S400 outperforms S200 on 6/9 tasks and achieves the best mean across the three resolutions on 5/9 tasks (Reply Table 6).

Table 6: Parcellation-resolution comparison

The best mean in each row is bolded.

Task / metricS100S200S400
HCP / Sex / weighted F1 ↑0.7551 ± 0.01580.7561 ± 0.02560.7875 ± 0.0115
ABCD / Sex / weighted F1 ↑0.6041 ± 0.02560.5793 ± 0.03330.5889 ± 0.0178
PPMI / Group / weighted F1 ↑0.5811 ± 0.02650.5844 ± 0.01700.5802 ± 0.0206
ADNI / AD / weighted F1 ↑0.6783 ± 0.07750.7246 ± 0.10440.7219 ± 0.0657
ADNI / MCI / weighted F1 ↑0.5498 ± 0.06600.5718 ± 0.05560.5461 ± 0.0553
ABIDE / ASD / weighted F1 ↑0.5300 ± 0.08510.5856 ± 0.04240.6013 ± 0.1107
ADHD200 / Diagnosis / weighted F1 ↑0.5711 ± 0.01820.5826 ± 0.02160.5872 ± 0.0215
ADHD200 / Age / nMAE ↓0.6759 ± 0.01770.6683 ± 0.04070.6523 ± 0.0315
SALD / Age / nMAE ↓0.7744 ± 0.04620.6999 ± 0.06730.6981 ± 0.0586

4. Prediction ablation and reliability (W4, Q4):

Thank you for highlighting the importance of the no-prediction control and the reliability of the results.

The original “No prediction” variant retains Decay-25. We added a control keeping consistency and SIGReg active throughout all 150 epochs (9,600 updates). The full model has better means on seven of nine tasks than this constant-consistency control. Keeping consistency active also improves the no-prediction variant on eight of nine tasks relative to Decay-25. Both settings are distinguished in the main-text objective-ablation table and the complete nine-task appendix table.

The full objective also has the best cohort-averaged rank (1.79) among five objective settings. A cohort-blocked permutation test detects overall differences among these settings (p = 0.0004). Together with the constant-consistency control, these results support the overall effectiveness of the complete training objective (Reply Table 7).

Table 7: Objective-ablation control (five-fold mean ± SD)

The table presents the objective settings discussed here; the complete objective ablation, including the no-consistency control, is reported in the manuscript appendix.

Task / metricFullNo pred: Decay-25No pred: ConstantNo SIGReg
HCP / Sex / weighted F1 ↑0.7874 ± 0.01150.6565 ± 0.01870.7474 ± 0.00710.6942 ± 0.0153
ABCD / Sex / weighted F1 ↑0.5899 ± 0.01840.5383 ± 0.01740.5747 ± 0.02620.5511 ± 0.0379
PPMI / Group / weighted F1 ↑0.5802 ± 0.02060.5399 ± 0.04600.5819 ± 0.04770.5426 ± 0.0720
ADNI / AD / weighted F1 ↑0.7219 ± 0.06570.5023 ± 0.08500.6302 ± 0.10600.6065 ± 0.0667
ADNI / MCI / weighted F1 ↑0.5461 ± 0.05530.4872 ± 0.09010.5401 ± 0.05950.5377 ± 0.0410
ABIDE / ASD / weighted F1 ↑0.6013 ± 0.11070.6082 ± 0.07490.5505 ± 0.06460.4136 ± 0.0587
ADHD200 / Diagnosis / weighted F1 ↑0.5872 ± 0.02150.5348 ± 0.03090.5623 ± 0.01610.5643 ± 0.0288
SALD / Age / nMAE ↓0.6982 ± 0.05860.7674 ± 0.05510.6868 ± 0.03710.7625 ± 0.0577
ADHD200 / Age / nMAE ↓0.6523 ± 0.03150.7416 ± 0.05670.6751 ± 0.03350.7132 ± 0.0482
Reviewer_oMXVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CoFlu-Brain, a dynamic functional connectivity (FC) pre-training framework based on the assumptions of cofluctuation: a LeJEPA-style pre-training on a disjoint set of windowed FC frames. The CoFlu-Brain is evaluated on eight datasets and nine tasks against four baseline models, and shows promising performance over baselines. While the work addresses an important topic in the field, some concerns on the technical contribution, extensiveness of the experiments can be further clarified.

Critical strengths
  • Addresses an important topic in the field: FC-based generative pre-training for dynamic FC representation learning
  • Performance improvements of the proposed method over the presented baseline are promising
Critical weaknesses
  • The technical contribution is rather incremental: the proposed CoFlu-Brain is basically a window-based dynamic FC generative pre-training method with LeJEPA objectives.
  • Baseline methods can be compared more extensively.
  • The consistency claim is not clearly supported by the ablation results.
Questions for authors
  • Technical contribution: please consider toning down the claims about cofluctuations that are not experimentally supported. (e.g., We hypothesize that retained frame sets sampled from the same segment provide different views of subject-specific cofluctuation structure.)
  • Please consider adding baseline methods if applicable: BrainJEPA, NeuroSTORM.
  • Please elaborate on the consistency claim given the minimal performance change after the ablation study.
  • Would the CoFlu-Brain pretraining scheme work if the windowed FC frames are given as ROI-Timeseries matrices? Please empirically or theoretically justify the advantage of windowed FC pre-training over the windowed ROI-Timeseries pre-training with the same retain-drop frame.
  • Minor: Please revise mathematical notations (e.g. define TT, NN, inconsistent use between zz (page 3) and ZZ (Figure 2))
3 comments

Thank you for the helpful comments. We address related points together below.

1. Contribution and wording (W1, Q1):

Thank you for suggesting a more precise description of the method.

We revised the Introduction and Methods to describe learning relationships among FC estimates from different frame compositions through retained-view consistency and dropped-view prediction. We also revised the “subject-specific cofluctuation structure” interpretation: consistency encourages different retained views of the same segment to form consistent connectivity representations.

2. Additional baselines (W2, Q2):

Thank you for recommending additional baseline comparisons.

We added Brain-JEPA (450 regions: Schaefer-400 plus Tian-50), Brain-DiT (424 regions), and an FC mean-readout baseline requiring no pretraining. All use the same downstream participant splits and linear-probe protocol; implementation details are in Appendix A. Table 1 results against the added baselines are shown below; classification uses weighted F1 and age prediction uses nMAE. Values are mean ± SD across five folds (Reply Table 1).

Table 1: Added-baseline comparison (manuscript Table 1; five-fold mean ± SD)

Task / metricCoFlu-BrainBrain-JEPA (450)Brain-DiT (424)FC mean
HCP / Sex / weighted F1 ↑0.7745 ± 0.01030.5766 ± 0.02070.7187 ± 0.01350.7304 ± 0.0243
ABCD / Sex / weighted F1 ↑0.5961 ± 0.01900.5189 ± 0.02020.5575 ± 0.03300.5797 ± 0.0295
PPMI / Group / weighted F1 ↑0.5775 ± 0.01730.5327 ± 0.02940.5194 ± 0.03300.5455 ± 0.0458
ADNI / AD / weighted F1 ↑0.6835 ± 0.03070.5589 ± 0.05730.6251 ± 0.06480.6137 ± 0.0255
ADNI / MCI / weighted F1 ↑0.5538 ± 0.04410.5010 ± 0.04750.5160 ± 0.04800.5238 ± 0.0309
ABIDE / ASD / weighted F1 ↑0.5567 ± 0.04760.5048 ± 0.04610.5884 ± 0.05040.5569 ± 0.0830
ADHD200 / Diagnosis / weighted F1 ↑0.5891 ± 0.01510.5231 ± 0.02860.5867 ± 0.02200.5397 ± 0.0210
SALD / Age / nMAE ↓0.6571 ± 0.04311.1583 ± 0.15000.8813 ± 0.04321.2716 ± 0.1386
ADHD200 / Age / nMAE ↓0.6557 ± 0.03700.9341 ± 0.06000.7794 ± 0.04380.7667 ± 0.0486

With these added baselines, CoFlu-Brain retains the best cohort-averaged rank, 1.43 among eight methods, and the best primary-metric mean on seven of nine tasks (Reply Table 2).

Table 2: Overall ranking across all eight methods

ModelCohort-averaged rank ↓
CoFlu-Brain1.43
FC mean readout3.93
BrainMASS4.00
Brain-DiT4.14
LCM4.43
BrainGFM4.64
BrainLM6.36
Brain-JEPA7.07

3. Consistency objective (W3, Q3):

We appreciate your question about how the ablations relate to the role of consistency.

Role of consistency.

Consistency encourages agreement across retained views early in training; its weight then decreases to emphasize dropped-view prediction. Retaining consistency improves mean performance on six of nine tasks in the objective ablation. The with/without-consistency comparison is in the manuscript appendix; Table 3 presents the remaining objective controls.

Table 3: Objective ablation (five-fold mean ± SD)

The table presents the objective settings discussed here; the complete objective ablation, including the no-consistency control, is reported in the manuscript appendix.

Task / metricFullNo pred: Decay-25No pred: ConstantNo SIGReg
HCP / Sex / weighted F1 ↑0.7874 ± 0.01150.6565 ± 0.01870.7474 ± 0.00710.6942 ± 0.0153
ABCD / Sex / weighted F1 ↑0.5899 ± 0.01840.5383 ± 0.01740.5747 ± 0.02620.5511 ± 0.0379
PPMI / Group / weighted F1 ↑0.5802 ± 0.02060.5399 ± 0.04600.5819 ± 0.04770.5426 ± 0.0720
ADNI / AD / weighted F1 ↑0.7219 ± 0.06570.5023 ± 0.08500.6302 ± 0.10600.6065 ± 0.0667
ADNI / MCI / weighted F1 ↑0.5461 ± 0.05530.4872 ± 0.09010.5401 ± 0.05950.5377 ± 0.0410
ABIDE / ASD / weighted F1 ↑0.6013 ± 0.11070.6082 ± 0.07490.5505 ± 0.06460.4136 ± 0.0587
ADHD200 / Diagnosis / weighted F1 ↑0.5872 ± 0.02150.5348 ± 0.03090.5623 ± 0.01610.5643 ± 0.0288
SALD / Age / nMAE ↓0.6982 ± 0.05860.7674 ± 0.05510.6868 ± 0.03710.7625 ± 0.0577
ADHD200 / Age / nMAE ↓0.6523 ± 0.03150.7416 ± 0.05670.6751 ± 0.03350.7132 ± 0.0482

Weight schedule.

The schedule ablation compares Decay-25, Decay-50, and constant weighting: Decay-25 has the best mean on five of seven classification tasks, whereas constant weighting performs best on ABIDE and both age tasks. These results support using consistency early in training and reducing its weight to balance consistency and prediction (main-text “Consistency-weight schedule” and complete appendix results) (Reply Table 4).

Table 4: Consistency-weight schedule (five-fold mean ± SD)

Task / metricDecay-25Decay-50Constant
HCP / Sex / weighted F1 ↑0.7737 ± 0.00920.7588 ± 0.02040.7687 ± 0.0164
ABCD / Sex / weighted F1 ↑0.5604 ± 0.01490.5955 ± 0.03490.5916 ± 0.0442
PPMI / Group / weighted F1 ↑0.5562 ± 0.05820.5179 ± 0.01200.5540 ± 0.0048
ADNI / AD / weighted F1 ↑0.7235 ± 0.03360.6736 ± 0.02360.6627 ± 0.0350
ADNI / MCI / weighted F1 ↑0.5967 ± 0.04220.5636 ± 0.01070.5431 ± 0.0253
ABIDE / ASD / weighted F1 ↑0.5427 ± 0.05000.4734 ± 0.02430.5669 ± 0.0179
ADHD200 / Diagnosis / weighted F1 ↑0.5756 ± 0.02980.5433 ± 0.02820.5688 ± 0.0300
SALD / Age / nMAE ↓0.7610 ± 0.08020.7607 ± 0.02610.7576 ± 0.0617
ADHD200 / Age / nMAE ↓0.6785 ± 0.02240.6839 ± 0.02390.6568 ± 0.0270

4. FC versus ROI time series (Q4):

Thank you for suggesting this input comparison.

We added a time-series pretraining control with the same data, random retain–drop sampling, four view pairs, consistency and latent-prediction objectives, and SIGReg. A temporal CNN produces one token per ROI, followed by the same six-layer, 512-dimensional Transformer and mean pooling. Both use the same participant-grouped five-fold linear-probe protocol.

Comparing the released FC model with the time-series control using one pretraining seed per variant, FC has better means on seven of nine tasks. FC versus time series: HCP sex weighted F1, 0.7875 versus 0.6665; ABCD sex, 0.5889 versus 0.5172; ADNI AD, 0.7219 versus 0.6216; ADHD age nMAE, 0.6523 versus 0.7278. These results provide empirical support for using FC representations in this framework (Reply Table 5).

Table 5: FC versus time-series input (mean scores; one pretraining seed per variant)

Task / metricReleased FCROI time series
HCP / Sex / weighted F1 ↑0.78750.6665
ABCD / Sex / weighted F1 ↑0.58890.5172
PPMI / Group / weighted F1 ↑0.58020.5544
ADNI / AD / weighted F1 ↑0.72190.6216
ADNI / MCI / weighted F1 ↑0.54610.5371
ABIDE / ASD / weighted F1 ↑0.60130.5808
ADHD200 / Diagnosis / weighted F1 ↑0.58720.6153
ADHD200 / Age / nMAE ↓0.65230.7278
SALD / Age / nMAE ↓0.69810.6919

5. Notation (Q5):

Thank you for pointing out the notation issues.

We clarified the notation definitions in the text and revised the corresponding Figure 2 labels.

I appreciate the authors’ response. Additional clarifications and experiments have addressed my questions. I am willing to keep my score, given that my concern about the level of contribution to the field remains (which I am also aware is independent of the soundness of the rebuttal response, which the authors did well). Still, I would not mind if the paper were accepted.

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

The authors propose CoFlu-Brain, a self-supervised foundation model over fMRI functional connectivity. They build several views from each scan by retaining and dropping frames, then combine a consistency loss on the retained-frame FC with a prediction loss on the latent representation of the complementary dropped frames. The model is pretrained on 13,444 subjects from 8 cohorts and evaluated with frozen encoders and linear probes on 9 downstream tasks, where it reports gains over several previously published fMRI representation models. Ablations are provided for random frame splitting, each loss component, data coverage, and the number of branches.

Critical strengths
  • Treating FC’s dependence on frame composition as a self-supervision signal is a clever and well-motivated idea.

  • Pretraining on 13,444 subjects from eight datasets.

  • Open source the code.

Critical weaknesses
  • The work is largely incremental. The proposed model is essentially a combination of existing techniques, and there is no fundamental architectural novelty.
Questions for authors

N/A

3 comments

Thank you for recognizing the motivation, pretraining scale, and open-source release, and for your comments on novelty.

1. Core innovation and its significance.

CoFlu-Brain starts from the dependence of FC on frame composition and constructs multiple connectivity views from different cofluctuation compositions within the same segment. Retained-view consistency and retained-to-complementary-dropped latent prediction encourage representations that capture connectivity information shared across temporal samples and predict information supported by unobserved frames. This design turns observational variation in FC estimation into self-supervision. The revised Introduction clarifies this contribution.

2. Comparison with direct FC readout.

We added an FC mean-readout baseline requiring no pretraining. It extracts 400-dimensional features from the same Schaefer-400 FC matrices and 200-frame segments, using identical downstream participant splits and fixed linear probes. CoFlu-Brain achieves better primary-metric means on eight of nine tasks, including HCP sex F1 of 0.7745 versus 0.7304, supporting the transfer gains from the learned representations (Reply Table 1).

Table 1: CoFlu-Brain versus FC mean readout (five-fold mean ± SD)

The best mean in each row is bolded.

Task / metricCoFlu-BrainFC mean readout
HCP / Sex / weighted F1 ↑0.7745 ± 0.01030.7304 ± 0.0243
ABCD / Sex / weighted F1 ↑0.5961 ± 0.01900.5797 ± 0.0295
PPMI / Group / weighted F1 ↑0.5775 ± 0.01730.5455 ± 0.0458
ADNI / AD / weighted F1 ↑0.6835 ± 0.03070.6137 ± 0.0255
ADNI / MCI / weighted F1 ↑0.5538 ± 0.04410.5238 ± 0.0309
ABIDE / ASD / weighted F1 ↑0.5567 ± 0.04760.5569 ± 0.0830
ADHD200 / Diagnosis / weighted F1 ↑0.5891 ± 0.01510.5397 ± 0.0210
SALD / Age / nMAE ↓0.6571 ± 0.04311.2716 ± 0.1386
ADHD200 / Age / nMAE ↓0.6557 ± 0.03700.7667 ± 0.0486

3. Distinction from BrainMASS and performance across resolutions.

BrainMASS uses frame dropping to construct FC augmentations for masked modeling and representation alignment. CoFlu-Brain additionally turns the corresponding dropped frames into prediction targets, jointly learning multi-view consistency and complementary prediction. Under the same downstream splits and linear-probe protocol, the Schaefer-100, -200, and -400 variants achieve better means than BrainMASS (S100) on seven, eight, and eight of nine tasks, respectively. At S100, ADNI AD F1 is 0.6783 versus 0.5811, also showing that the gains are not solely attributable to finer parcellation (Reply Table 2).

Table 2: Comparison with BrainMASS across Schaefer resolutions (five-fold mean ± SD)

Bold values indicate CoFlu-Brain means better than the BrainMASS (S100) reference on the same task, rather than the best value in each row.

Task / metricBrainMASS (S100)CoFlu-Brain (S100)CoFlu-Brain (S200)CoFlu-Brain (S400)
HCP / Sex / weighted F1 ↑0.7203 ± 0.00480.7551 ± 0.01580.7561 ± 0.02560.7875 ± 0.0115
ABCD / Sex / weighted F1 ↑0.6215 ± 0.05150.6041 ± 0.02560.5793 ± 0.03330.5889 ± 0.0178
PPMI / Group / weighted F1 ↑0.5751 ± 0.03010.5811 ± 0.02650.5844 ± 0.01700.5802 ± 0.0206
ADNI / AD / weighted F1 ↑0.5811 ± 0.03930.6783 ± 0.07750.7246 ± 0.10440.7219 ± 0.0657
ADNI / MCI / weighted F1 ↑0.5045 ± 0.03460.5498 ± 0.06600.5718 ± 0.05560.5461 ± 0.0553
ABIDE / ASD / weighted F1 ↑0.5472 ± 0.05270.5300 ± 0.08510.5856 ± 0.04240.6013 ± 0.1107
ADHD200 / Diagnosis / weighted F1 ↑0.5549 ± 0.02540.5711 ± 0.01820.5826 ± 0.02160.5872 ± 0.0215
ADHD200 / Age / nMAE ↓0.9328 ± 0.03770.6759 ± 0.01770.6683 ± 0.04070.6523 ± 0.0315
SALD / Age / nMAE ↓0.9064 ± 0.08770.7744 ± 0.04620.6999 ± 0.06730.6981 ± 0.0586

With frame counts, data, architecture, losses, and training budget matched in the pairing control, complementary pairing outperforms independent sampling on six of nine tasks. The full objective outperforms a no-prediction control with consistency active throughout training on seven tasks. These experiments support the roles of pairing and prediction, respectively; complete results are reported in Appendix D and the objective-ablation appendix.

4. Functional-network interpretation.

Manuscript Figure 3 examines ADHD200 age and diagnosis prediction by retaining subject-specific FC within each Yeo-17 network and replacing other connections with probe-training means. Network rankings are consistent across three independently pretrained models, with mean pairwise rank correlations of 0.906 and 0.857 for the two tasks, respectively. Default-network components are prominent in both tasks, with the salience/ventral-attention system also contributing to diagnosis. This provides a functional-network interpretation of the learned representations.

OK, you don't need to include these. I'm not rejecting your paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.