Pith. sign in

REVIEW 4 major objections 5 minor 46 references

Brain Network Analysis Based on Fine-tuned Self-supervised Model for Brain Disease Diagnosis

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A small adapter on a frozen self-supervised brain-network transformer improves Alzheimer's classification on fMRI.

desk verdict Plausible adapter-on-brain-network idea, but the missing train/test split and a likely InfoNCE typo mean the reported superiority is not yet demonstrated. read the letter →

arxiv 2506.11671 v2 pith:7LDEC5EY submitted 2025-06-13 eess.IV cs.CV

Yifei Tang , Hongjie Jiang , Changhong Jing , Hieu Pham , Shuqiang Wang This is my paper
classification eess.IVcs.CV
keywords brainnetworkfunctionalMRIself-supervisedlearningtransformeradapterAlzheimer'sdiseasemildcognitiveimpairmentdiagnosis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a small trainable adapter placed in front of a frozen, self-supervised transformer encoder for brain networks improves the diagnosis of Alzheimer's disease and mild cognitive impairment from resting-state fMRI. On the ADNI dataset, the proposed model reports 78.35% accuracy for AD versus healthy controls and 64.96% accuracy for MCI versus healthy controls, outperforming two established baselines, BrainNetCNN and BrainGNN. The central idea is to expand the region-of-interest feature dimension through linear projections before feeding the functional connectivity matrix into the pre-trained encoder, then use the encoder's latent representation with an SVM classifier. If this holds, it would mean that a large pre-trained brain-network model can be adapted to a diagnostic task with only a tiny amount of trainable parameters, without modifying the foundation model.

What carries the argument

The central mechanism is the Adapter module, a feed-forward block with two linear projections and a LeakyReLU, that maps the $V \times V$ functional connectivity matrix to a $V \times B$ representation before entering the frozen BrainTF transformer encoder. The transformer then applies multi-head self-attention across the $V$ brain regions to capture inter-region associations. The model is trained with a weighted sum of an InfoNCE loss (classification) and an MSE loss (reconstruction), with balance parameters $\lambda_c$ and $\lambda_r$; these two losses together let the adapter expand the feature dimension while preserving the connectome structure.

What would settle it

Run the same ADNI experiment with a strict subject-disjoint split (e.g., 80/20 train/test) and check whether the 78.35% AD-versus-NC accuracy survives; if it drops to near-chance or to the level of the baseline models, the claimed gain is an artifact of double-dipping rather than a genuine diagnostic improvement.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a pre-trained brain-network transformer (BrainTF), trained self-supervised on fMRI from thousands of subjects, can be repurposed for disease classification by adding a two-layer adapter that maps the functional connectivity matrix from $V \times V$ to $V \times B$, while the encoder's weights stay frozen. The adapted representation is read out and fed to an SVM, yielding the reported accuracies. The paper also shows, through a reconstruction head and an InfoNCE classification head, that the model can both rebuild the connectivity matrix and support diagnosis, and that both loss terms contribute to the final performance.

Load-bearing premise

The reported accuracies are only meaningful if the subjects used to fit the adapter and the SVM were not also used to compute the accuracy; the paper does not describe a train/validation/test split, so if the same subjects appear in both phases, the numbers are in-sample fits rather than predictions.

Editorial extensions

If this is right

  • A pre-trained brain-network transformer can be adapted to a new diagnostic task by training only a small adapter, avoiding the need to train the full encoder on labeled data.
  • The reported 78.35% AD versus NC accuracy is roughly 7 to 11 percentage points higher than the two baselines, suggesting the adapter-plus-frozen-encoder combination extracts more discriminative features.
  • The reconstruction head can reproduce the functional connectivity matrix, which the paper uses as visual evidence that the model captures disease-relevant inter-regional relationships.
  • The ablation shows that both the InfoNCE loss and the MSE loss matter: removing either lowers accuracy, particularly the classification loss in the MCI versus NC task.
  • The approach keeps the foundation model frozen, so the same pre-trained weights could be shared across many downstream diagnostic tasks with only small task-specific adapters.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the accuracy holds under a strict split, the adapter design suggests a general recipe for medical-imaging foundation models: keep the large encoder frozen and learn a small projection per task, which could make multi-task deployment much cheaper.
  • The relatively low MCI-versus-NC accuracy (64.96%) hints that static ROI correlation matrices derived from a single scan may not carry enough signal for early-stage disease; combining the adapter with temporal dynamics or multimodal inputs might be a natural next test.
  • A direct comparison against fine-tuning the transformer itself, or against a linear probe on the frozen features, would isolate whether the adapter's dimensional expansion is the actual source of the gain rather than the frozen features alone.
  • The reconstruction head offers a testable extension: if the reconstructed connectomes faithfully preserve disease-related edges, they could serve as synthetic training data for other models in low-data settings.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes a parameter-efficient adapter for brain-network analysis. It maps a 90x90 functional connectivity matrix into a higher-dimensional representation, feeds it into a frozen self-supervised transformer encoder (referred to as BrainTF), and uses the resulting latent representation with an SVM for Alzheimer's disease (AD) vs normal control (NC) and mild cognitive impairment (MCI) vs NC classification on ADNI. The Adapter and two auxiliary heads are trained with a weighted sum of a reconstruction (MSE) loss and an InfoNCE classification loss. The authors report that the proposed method achieves 78.35% accuracy for AD vs NC and 64.96% for MCI vs NC, outperforming BrainNetCNN and BrainGNN, and they provide an ablation study on the loss weights. The core claim is that a lightweight adapter on a frozen brain-network foundation model improves diagnostic performance.

Significance. If the reported results are genuine out-of-sample predictions, the paper makes a useful contribution: it shows that a small trainable adapter on top of a frozen, self-supervised brain-network foundation model can improve downstream diagnosis relative to task-specific CNN/GNN baselines. The idea is simple, the experimental target is clinically relevant, and the use of a large-scale pretrained encoder is timely. The paper is clearly written and the proposed architecture is easy to understand. However, the evidence as presented is not yet sufficient to support the central claim, because the evaluation protocol does not establish that the reported accuracies are predictions on held-out subjects, and Eq. (9) as written is a constant. The missing out-of-sample protocol and the missing direct BrainTF-without-adapter baseline are the decisive issues; both are fixable in revision.

major comments (4)
  1. [4.1] Section 4.1 never describes a train/validation/test split. The dataset is described only as containing 64 AD, 135 MCI, and 263 NC subjects, and the implementation details mention the optimizer, loss weights, and training epochs, but not how subjects are partitioned. Since the Adapter, both heads, and the SVM are fit on the same data used to compute Tables 1 and 2, the reported accuracies may be in-sample training scores rather than out-of-sample predictions. The statement that the experiment is repeated three times and averaged does not create a held-out evaluation. Please specify the exact subject-level split (or nested cross-validation) and report all metrics on a held-out partition; this is load-bearing for the abstract's claim of superior diagnostic performance.
  2. [3.4, Eq. (9)] Equation (9) is written with the same term sim(q,k+) in both the numerator and every term of the denominator, so the ratio is 1/N and Lc = log N, a constant with no gradient. This cannot provide any training signal for the classification head. In contrast, Table 3 shows that λc changes accuracy, which implies that the implemented loss is different from what is shown. Please correct Eq. (9) to the standard InfoNCE form with negative samples, and ensure the equation matches the actual implementation.
  3. [4.2, Tables 1-2] The comparison treats BrainNetCNN and BrainGNN as the only baselines. Because the proposed method is a frozen foundation model plus an adapter, the natural control is the same frozen BrainTF with a readout or SVM but without the adapter. Without this baseline, the reported improvement cannot be attributed to the adapter; it may simply reflect the pretrained self-supervised features of BrainTF. Please add this control, and ideally also a linear-probe result on the frozen encoder, to isolate the adapter's contribution.
  4. [Tables 1-3] All reported metrics are single averages over three repetitions, with no standard deviations, per-class details, or statistical significance tests. The claimed 7-11% accuracy improvements over baselines cannot be evaluated for reliability. At a minimum, please report mean ± standard deviation over repeated runs with different seeds or cross-validation folds, and state whether the same folds are used for all compared methods.
minor comments (5)
  1. [3.1] The paper calls the model 'fine-tuned self-supervised', but Section 3.1 says the parameters of BrainTF are frozen and only the Adapter is trained. This is adapter-based probing rather than fine-tuning; please align the terminology with the actual training procedure.
  2. [3.3, Eqs. (2)-(4)] Equations (2)-(4) use the same weight matrix W_q for the query, key, and value projections. The key and value should presumably use separate matrices W_k and W_v; this appears to be a typographical error.
  3. [Fig. 1] The labels B, S, M, and V in Figure 1 are not defined in the text or caption, and the adapter output dimension B from Section 3.2 is not given a numerical value in Section 4.1. Please define all symbols used in the figure.
  4. [4.1] The paper does not state how the 90x90 functional connectivity is computed from the BOLD time-series (e.g., full Pearson correlation, partial correlation, or number of time points), nor the exact ADNI preprocessing steps. This information is needed for reproducibility.
  5. [4.4, Table 3] The checkmark format in Table 3 is ambiguous; each row should list the actual numerical values of λc and λr, rather than checkmarks.

Circularity Check

1 steps flagged · score 5.0 of 10

Section 4.1 reports no train/validation/test split, so the headline accuracies are indistinguishable from in-sample fits rather than diagnostic predictions.

  1. fitted input called prediction [Section 4.1 (Datasets and Experiment Design); Tables 1 and 2]
    "Datasets.In this experiment, we use the Alzheimer's Disease Neuroimaging Initiative(ADNI) datasets [44], which contains 64 Alzheimer's disease(AD), 135 mild cognitive impairment(MCI) group and 263 normal controls(NC). ... In the downstream tasks, we apply the SVM classifier for prediction. To ensure the robustness of our results, we repeat the experiment three times and take the average."

    The paper reports the AD vs NC and MCI vs NC accuracies as the downstream 'prediction' of the frozen BrainTF + Adapter + SVM pipeline, but it never states that the 64 AD, 135 MCI, and 263 NC subjects were partitioned into training and held-out test sets. The only evaluation procedure described is training the adapter and SVM on the ADNI data and repeating the experiment three times; repeating on the same data does not create out-of-sample predictions. As written, Tables 1 and 2 are therefore consistent with being in-sample training scores. If the same subjects were used for fitting and reporting, the claimed 'superior performance' in the abstract reduces to fitting accuracy, not to a diagnostic prediction, which is the fitted-input-called-prediction pattern.

full rationale

The central claim is that the frozen BrainTF + Adapter + SVM classifier achieves 78.35% (AD vs NC) and 64.96% (MCI vs NC) and outperforms BrainNetCNN and BrainGNN. A diagnostic prediction requires the adapter and SVM to be evaluated on subjects not used to fit them. Section 4.1 names the ADNI subject counts, the optimizer, the loss weights, and the SVM, but never describes a train/validation/test split; 'we repeat the experiment three times and take the average' is a repetition on the same data, not cross-validation. If the same 64/135/263 subjects are used for fitting and reporting, Tables 1 and 2 are in-sample fits, and the abstract's superiority claim is a fitted-input-called-prediction. This is the load-bearing issue. I do not find a load-bearing self-citation chain: the many same-group citations are background, and BrainTF's pretraining is asserted rather than derived from a cited uniqueness theorem. Separately, Eq. 9 as written has the same term sim(q,k+) in the numerator and denominator, making Lc constant; that is an internal consistency/typo problem, not a circularity. The score reflects partial circularity in the evaluation, not a proof of in-sample fitting, because the text is silent rather than explicit about the split.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim depends on a handful of hand-set hyperparameters (loss weights, unspecified temperature and output dimension) and on the untested transferability of BrainTF. The paper introduces no new physical entities.

free parameters (4)
  • lambda_c (classification loss weight) = 0.2
    Hand-set weighting of the InfoNCE loss in Eq. 10; the ablation in Table 3 shows it materially affects MCI vs NC accuracy.
  • lambda_r (reconstruction loss weight) = 5
    Hand-set weighting of MSE loss in Eq. 10; its removal degrades performance.
  • InfoNCE temperature tau = not specified
    Temperature in Eq. 9 is required for the loss but no value is reported.
  • Adapter output dimension B = not specified
    The core dimension-expansion hyperparameter of the Adapter is never given a value.
assumptions (3)
  • domain assumption Pearson correlation of BOLD time series yields a meaningful functional connectivity matrix.
    Used to construct input R^{V x V} in Section 3.1; a standard modeling choice in brain-network analysis.
  • domain assumption BrainTF, the pre-trained foundation model from BrainMass [17], has weights that transfer to the ADNI dataset.
    The paper freezes BrainTF and relies entirely on its pre-trained representations (Section 3.3) without fine-tuning or verification.
  • ad hoc to paper The InfoNCE loss as written in Eq. 9 is a valid classification objective.
    The equation is degenerate because the denominator uses the positive sample k+ for every term, making the loss identically zero; the paper does not acknowledge this.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Brain Network Analysis Based on Fine-tuned Self-supervised Model for Brain Disease Diagnosis." pith.science (2026). https://pith.science/paper/7LDEC5EY

@misc{pith2026250611671,
  author       = {Pith},
  title        = {Pith review of: Brain Network Analysis Based on Fine-tuned Self-supervised Model for Brain Disease Diagnosis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7LDEC5EY}},
  note         = {Machine review of arXiv:2506.11671}
}
read the original abstract

Functional brain network analysis has become an indispensable tool for brain disease analysis. It is profoundly impacted by deep learning methods, which can characterize complex connections between ROIs. However, the research on foundation models of brain network is limited and constrained to a single dimension, which restricts their extensive application in neuroscience. In this study, we propose a fine-tuned brain network model for brain disease diagnosis. It expands brain region representations across multiple dimensions based on the original brain network model, thereby enhancing its generalizability. Our model consists of two key modules: (1)an adapter module that expands brain region features across different dimensions. (2)a fine-tuned foundation brain network model, based on self-supervised learning and pre-trained on fMRI data from thousands of participants. Specifically, its transformer block is able to effectively extract brain region features and compute the inter-region associations. Moreover, we derive a compact latent representation of the brain network for brain disease diagnosis. Our downstream experiments in this study demonstrate that the proposed model achieves superior performance in brain disease diagnosis, which potentially offers a promising approach in brain network analysis research.

Figures

Figures reproduced from arXiv: 2506.11671 by the authors.

Figure 1
Figure 1. The framework of our method is divided into three parts. I.Construction of FC (Functional Connectivity). II.Training module which composed of Adapter and Brain network Transformer encoder. III.Testing module is for downstream tasks. The dark arrow means the overall data flow. Green dashed line indicates loss calculation [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Radar plot about our evaluation of two baseline models. (a) shows the metrics in classification task AD vs NC. It is obvious that our model perform best in four metrics. (b)shows the metrics in classification task MCI vs NC. Our model perform best in acc, spe and F1-score [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Input and Output brain network from AD patient. The brain network con￾structed by our model can effectively capture the inter-regional relationships of the brain in Alzheimer’s disease patients, providing strong evidence to support our perfor￾mance in downstream tasks. 4.4 Ablation Study In Eq. 10, we apply two parameters to calculate loss. In ablation study, we aim to discover whether parameters λc and λr necessary… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 29 canonical work pages

  1. [1]

    Nature 453(7197), 869–878 (2008)

    Logothetis, N.K.: What we can do and what we cannot do with fmri. Nature 453(7197), 869–878 (2008)

    work page 2008

  2. [2]

    Heeger, D.J., Ress, D.: What does fmri tell us about neuronal activity? Nature reviews neuroscience3(2), 142–151 (2002)

    work page 2002

  3. [3]

    nature412(6843), 150–157 (2001)

    Logothetis, N.K., Pauls, J., Augath, M., Trinath, T., Oeltermann, A.: Neurophys- iological investigation of the basis of the fmri signal. nature412(6843), 150–157 (2001)

    work page 2001

  4. [4]

    https://doi.org/https://doi.org/10.1016/j.neubiorev.2004.10.009

    Fingelkurts, A.A., Fingelkurts, A.A., K¨ ahk¨ onen, S.: Functional connectivity in the brain—is it an elusive concept? Neuroscience & Biobehavioral Reviews28(8), 827– 836 (2005). https://doi.org/https://doi.org/10.1016/j.neubiorev.2004.10.009

    doi:10.1016/j.neubiorev.2004.10.009 2005

  5. [5]

    European Neuropsychopharmacology20(8), 519–534 (2010)

    van den Heuvel, M.P., Hulshoff Pol, H.E.: Exploring the brain network: A review on resting-state fmri functional connectiv- ity. European Neuropsychopharmacology20(8), 519–534 (2010). https://doi.org/https://doi.org/10.1016/j.euroneuro.2010.03.008

    doi:10.1016/j.euroneuro.2010.03.008 2010

  6. [6]

    Proceedings of the National Academy of Sciences101(13), 4637–4642 (2004)

    Greicius, M.D., Srivastava, G., Reiss, A.L., Menon, V.: Default-mode network ac- tivity distinguishes alzheimer’s disease from healthy aging: evidence from func- tional mri. Proceedings of the National Academy of Sciences101(13), 4637–4642 (2004)

    work page 2004

  7. [7]

    Human brain mapping26(4), 231–239 (2005)

    Rombouts, S.A., Barkhof, F., Goekoop, R., Stam, C.J., Scheltens, P.: Altered rest- ing state networks in mild cognitive impairment and mild alzheimer’s disease: an fmri study. Human brain mapping26(4), 231–239 (2005)

    work page 2005

  8. [8]

    Biological psychiatry62(5), 429–437 (2007) Title Suppressed Due to Excessive Length 11

    Greicius, M.D., Flores, B.H., Menon, V., Glover, G.H., Solvason, H.B., Kenna, H., Reiss, A.L., Schatzberg, A.F.: Resting-state functional connectivity in major depression: abnormally increased contributions from subgenual cingulate cortex and thalamus. Biological psychiatry62(5), 429–437 (2007) Title Suppressed Due to Excessive Length 11

    work page 2007

Show all 46 references
  1. [9]

    IEEE Trans- actions on Consumer Electronics (2024)

    Chen, X., Yao, W., Li, Y., Liang, D., Zheng, H., Shakil, S., Wang, S., Sun, T.: Ig- gcn: Empowering e-health services for alzheimer’s disease prediction. IEEE Trans- actions on Consumer Electronics (2024)

    2024

  2. [10]

    IEEE Transactions on Circuits and Systems for Video Technology (2025)

    Guo, X., Chen, X., Wang, S., Pun, C.M.: Underwater image restoration through a prior guided hybrid sense approach and extensive benchmark analysis. IEEE Transactions on Circuits and Systems for Video Technology (2025)

    2025

  3. [11]

    NeuroImage146, 1038– 1049 (2017)

    Kawahara, J., Brown, C.J., Miller, S.P., Booth, B.G., Chau, V., Grunau, R.E., Zwicker, J.G., Hamarneh, G.: Brainnetcnn: Convolutional neural networks for brain networks; towards predicting neurodevelopment. NeuroImage146, 1038– 1049 (2017)

    2017

  4. [12]

    Medical Im- age Analysis83, 102679 (2023)

    Huang, J., Wang, M., Ju, H., Shi, Z., Ding, W., Zhang, D.: Sd-cnn: A static- dynamic convolutional neural network for functional brain networks. Medical Im- age Analysis83, 102679 (2023)

    2023

  5. [13]

    Deep learning applications pp

    Wang, S., Wang, H., Cheung, A.C., Shen, Y., Gan, M.: Ensemble of 3d densely connected convolutional network for diagnosis of mild cognitive impairment and alzheimer’s disease. Deep learning applications pp. 53–73 (2020)

    2020

  6. [14]

    Medical Image Analysis74, 102233 (2021)

    Li, X., Zhou, Y., Dvornek, N., Zhang, M., Gao, S., Zhuang, J., Scheinost, D., Staib, L.H., Ventola, P., Duncan, J.S.: Braingnn: Interpretable brain graph neural network for fmri analysis. Medical Image Analysis74, 102233 (2021)

    2021

  7. [15]

    Scientific reports11(1), 8061 (2021)

    Wein, S., Malloni, W.M., Tom´ e, A.M., Frank, S.M., Henze, G.I., W¨ ust, S., Greenlee, M.W., Lang, E.W.: A graph neural network framework for causal inference in brain networks. Scientific reports11(1), 8061 (2021)

    2021

  8. [16]

    Advances in Neural Information Processing Systems35, 25586–25599 (2022)

    Kan, X., Dai, W., Cui, H., Zhang, Z., Guo, Y., Yang, C.: Brain network trans- former. Advances in Neural Information Processing Systems35, 25586–25599 (2022)

    2022

  9. [17]

    IEEE Transactions on Medical Imaging (2024)

    Yang, Y., Ye, C., Su, G., Zhang, Z., Chang, Z., Chen, H., Chan, P., Yu, Y., Ma, T.: Brainmass: Advancing brain network analysis for diagnosis with large-scale self-supervised learning. IEEE Transactions on Medical Imaging (2024)

    2024

  10. [18]

    Wang, S., Yanyan, S., Zhang, W.: Enhanced generative adversarial network and target sample recognition method (Nov 26 2024), uS Patent 12,154,036

    2024

  11. [19]

    International Journal of Nonlinear Sciences and Numerical Simulation11(8), 625–630 (2010)

    Liu, Y., Ren, Z.F., Wang, S.Q.: Theoretical analysis and experimental verification of the effect of surface tension on morphology of nanofibers in bubble electrospin- ning. International Journal of Nonlinear Sciences and Numerical Simulation11(8), 625–630 (2010)

    2010

  12. [20]

    IEEE Journal of Biomedical and Health Informatics27(8), 4154–4165 (2023)

    Wen, G., Cao, P., Liu, L., Yang, J., Zhang, X., Wang, F., Zaiane, O.R.: Graph self-supervised learning with application to brain networks analysis. IEEE Journal of Biomedical and Health Informatics27(8), 4154–4165 (2023)

    2023

  13. [21]

    In: 2024 Inter- national Joint Conference on Neural Networks (IJCNN)

    Chen, W., Lei, Y., Luo, S., Zhou, Z., Li, M., Pun, C.M.: Uwformer: Underwater image enhancement via a semi-supervised multi-scale transformer. In: 2024 Inter- national Joint Conference on Neural Networks (IJCNN). pp. 1–8. IEEE (2024)

    2024

  14. [22]

    In: International Conference on Neural Computing for Advanced Applications

    Yao, W., Shen, Y., Nicolls, F., Wang, S.Q.: Conditional diffusion model-based data augmentation for alzheimer’s prediction. In: International Conference on Neural Computing for Advanced Applications. pp. 33–46. Springer Nature Singapore Sin- gapore (2023)

    2023

  15. [23]

    In: Chinese Conference on Biometric Recognition

    Li, Y., Liao, I.Y., Zhong, N., Toshihiro, F., Wang, Y., Wang, S.: Generative ai en- ables the detection of autism using eeg signals. In: Chinese Conference on Biometric Recognition. pp. 375–384. Springer Nature Singapore Singapore (2023)

    2023

  16. [24]

    IEEE Transactions on Medical Imaging, DOI: 10.1109/TMI.2025.35502061(1), 1 (2025) 12 Y

    Yao, W., Lyu, Z., Mahmud, M., Zhong, N., Lei, B., Wang., S.: Catd: Unified representation learning for eeg-to-fmri cross-modal generation. IEEE Transactions on Medical Imaging, DOI: 10.1109/TMI.2025.35502061(1), 1 (2025) 12 Y. Tang et al

    2025

  17. [25]

    IEEE Trans- actions on Pattern Analysis and Machine Intelligence (2024)

    Gui, J., Chen, T., Zhang, J., Cao, Q., Sun, Z., Luo, H., Tao, D.: A survey on self-supervised learning: Algorithms, applications, and future trends. IEEE Trans- actions on Pattern Analysis and Machine Intelligence (2024)

    2024

  18. [26]

    In: Proceedings of the 2022 14th International Conference on Machine Learning and Computing

    You, S., Shen, Y., Wu, G., Wang, S.: Brain mr images super-resolution with the consistent features. In: Proceedings of the 2022 14th International Conference on Machine Learning and Computing. pp. 501–506 (2022)

    2022

  19. [27]

    Journal of Intelligent & Fuzzy Systems38(4), 3557–3566 (2020)

    Wang, S., Liang, S., Peng, F.: Image edge detection algorithm based on fuzzy set. Journal of Intelligent & Fuzzy Systems38(4), 3557–3566 (2020)

    2020

  20. [28]

    IEEE Transactions on Automation Science and Engineering (2024)

    Zuo, Q., Chen, L., Shen, Y., Ng, M.K.P., Lei, B., Wang, S.: Bdht: generative ai enables causality analysis for mild cognitive impairment. IEEE Transactions on Automation Science and Engineering (2024)

    2024

  21. [29]

    Advances in neural in- formation processing systems27(2014)

    Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial nets. Advances in neural in- formation processing systems27(2014)

    2014

  22. [30]

    In: Chinese Conference on Pattern Recognition and Computer Vision (PRCV)

    Kong, H., Pan, J., Shen, Y., Wang, S.: Adversarial learning based structural brain- network generative model for analyzing mild cognitive impairment. In: Chinese Conference on Pattern Recognition and Computer Vision (PRCV). pp. 361–375. Springer Nature Switzerland Cham (2022)

    2022

  23. [31]

    IEEE Transactions on Cybernetics54(9), 5026–5039 (2024)

    Jing, C., Shen, Y., Zhao, S., Pan, Y., Chen, C.P., Lei, B., Wang, S.: Estimating addiction-related brain connectivity by prior-embedding graph generative adver- sarial networks. IEEE Transactions on Cybernetics54(9), 5026–5039 (2024)

    2024

  24. [32]

    Advances in neural information processing systems33, 6840–6851 (2020)

    Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. Advances in neural information processing systems33, 6840–6851 (2020)

    2020

  25. [33]

    Brain Informatics11(1), 1 (2024)

    Jing, C., Kuai, H., Matsumoto, H., Yamaguchi, T., Liao, I.Y., Wang, S.: Addiction- related brain networks identification via graph diffusion reconstruction network. Brain Informatics11(1), 1 (2024)

    2024

  26. [34]

    Wang, S., Yu, W., Chenchen, X., Shengye, H.: Visualization method for evaluating brain addiction traits, apparatus, and medium (Sep 17 2024), uS Patent 12,093,833

    2024

  27. [35]

    Brain Informatics 10(1), 2 (2023)

    Gong, C., Chen, X., Mughal, B., Wang, S.: Addictive brain-network identification by spatial attention recurrent network with feature selection. Brain Informatics 10(1), 2 (2023)

    2023

  28. [36]

    IEEE transactions on neural networks20(1), 61–80 (2008)

    Scarselli, F., Gori, M., Tsoi, A.C., Hagenbuchner, M., Monfardini, G.: The graph neural network model. IEEE transactions on neural networks20(1), 61–80 (2008)

    2008

  29. [37]

    Computers in Biology and Medicine127, 104096 (2020)

    Jiang, H., Cao, P., Xu, M., Yang, J., Zaiane, O.: Hi-gcn: A hierarchical graph convolution network for graph embedding learning of brain network and brain disorders prediction. Computers in Biology and Medicine127, 104096 (2020)

    2020

  30. [38]

    Advances in neural information pro- cessing systems30(2017)

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need. Advances in neural information pro- cessing systems30(2017)

    2017

  31. [39]

    In: Interna- tional Conference on Neural Computing for Advanced Applications

    Dong, Y., Chen, X., Shen, Y., Ng, M.K.P., Qian, T., Wang, S.: Multi-modal mood reader: Pre-trained model empowers cross-subject emotion recognition. In: Interna- tional Conference on Neural Computing for Advanced Applications. pp. 178–192. Springer Nature Singapore Singapore (2024)

    2024

  32. [40]

    arXiv preprint arXiv:2108.05542 (2021)

    Kalyan, K.S., Rajasekharan, A., Sangeetha, S.: Ammus: A survey of transformer- based pretrained models in natural language processing. arXiv preprint arXiv:2108.05542 (2021)

    2021 arXiv

  33. [41]

    ACM computing surveys (CSUR)54(10s), 1–41 (2022)

    Khan, S., Naseer, M., Hayat, M., Zamir, S.W., Khan, F.S., Shah, M.: Transformers in vision: A survey. ACM computing surveys (CSUR)54(10s), 1–41 (2022)

    2022

  34. [42]

    Medical image analysis91, 102996 (2024) Title Suppressed Due to Excessive Length 13

    Zhang, S., Metaxas, D.: On the challenges and perspectives of foundation models for medical image analysis. Medical image analysis91, 102996 (2024) Title Suppressed Due to Excessive Length 13

    2024

  35. [43]

    arXiv preprint arXiv:1807.03748 (2018)

    Oord, A.v.d., Li, Y., Vinyals, O.: Representation learning with contrastive predic- tive coding. arXiv preprint arXiv:1807.03748 (2018)

    2018 arXiv

  36. [44]

    Neuroimaging Clinics15(4), 869–877 (2005)

    Mueller, S.G., Weiner, M.W., Thal, L.J., Petersen, R.C., Jack, C., Jagust, W., Trojanowski, J.Q., Toga, A.W., Beckett, L.: The alzheimer’s disease neuroimaging initiative. Neuroimaging Clinics15(4), 869–877 (2005)

    2005

  37. [45]

    Neuroimage15(1), 273–289 (2002)

    Tzourio-Mazoyer, N., Landeau, B., Papathanassiou, D., Crivello, F., Etard, O., Delcroix, N., Mazoyer, B., Joliot, M.: Automated anatomical labeling of activations in spm using a macroscopic anatomical parcellation of the mni mri single-subject brain. Neuroimage15(1), 273–289 (2002)

    2002

  38. [46]

    arXiv preprint arXiv:1412.6980 (2014)

    Kingma, D.P., Ba, J.: Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.