Pith. sign in

REVIEW 4 major objections 4 minor 28 references

SFC-GAN: A Generative Adversarial Network for Brain Functional and Structural Connectome Translation

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A CycleGAN extended with a structure-preserving loss can translate brain structural and functional connectivity in both directions while keeping network topology close to ground truth.

desk verdict A legitimate first attempt at bidirectional SC/FC translation, but the central claim of outperforming baselines is unverifiable because no baselines are reported, and the classification results contradict the paper's own conclusion. read the letter →

arxiv 2501.07055 v1 pith:KTIALLJA submitted 2025-01-13 cs.CV cs.LG

Yee-Fan Tan , Jun Lin Liow , Pei-Sze Tan , Fuad Noman , Raphael C.-W. Phan , Hernando Ombao , Chee-Ming Ting This is my paper · ORCID
classification cs.CVcs.LG
keywords brainconnectomestructuralconnectivityfunctionalbidirectionaltranslationCycleGANstructure-preservinglossgenerativeadversarialnetworktopology
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 introduces SFC-GAN, a generative adversarial network that translates between a brain's structural connectivity (SC, from diffusion MRI) and functional connectivity (FC, from fMRI) in both directions. Its central claim is that one consistent model can synthesize the missing connectome from whichever modality is available, while keeping the network's topology close to the ground truth. On the ADNI and DUMC-MDD datasets, the translated connectomes show lower reconstruction error and closer graph properties with the structure-preserving loss than without it, and they remain informative enough for SVM classification of Alzheimer's disease, mild cognitive impairment, and major depressive disorder. If the claim holds, studies with only one imaging modality can still obtain both connectomes, enabling joint structure-function analyses.

What carries the argument

The central mechanism is the cycle-consistent adversarial loop: a generator from SC to FC paired with a generator from FC to SC, so that translating a connectome and translating it back should recover the original. Around that loop, the structure-preserving loss (global and per-row Pearson correlation plus MSE) and a final symmetrization layer carry the topology-preservation claim, while the convolutional encoder-decoder stacks are the components claimed to capture the spatial structure of the 116 by 116 connectome matrices.

What would settle it

Randomly permute the 116 AAL region labels in every connectome, retrain SFC-GAN on the same data, and compare translation quality to the original ordering; if error and graph-property metrics are nearly identical, then the convolutional spatial-structure assumption plays no essential role.

Watch

Extended reading notes

Core claim

SFC-GAN is a CycleGAN with two generators, one mapping FC to SC and the other mapping SC to FC, plus two discriminators that judge whether translated connectomes look real. The training objective combines adversarial loss, L1 cycle-consistency loss, identity loss, and a structure-preserving loss made of global Pearson correlation, per-region (row-wise) Pearson correlation, and mean squared error between translated and true connectomes. A symmetrization layer forces the generator output to be a symmetric matrix, respecting the fact that a connectome is an undirected network. The paper reports that on ADNI and DUMC-MDD, the structure-preserving loss lowers MSE and MAE and aligns graph properties such as density, characteristic path length, global efficiency, and modularity more closely with ground truth, and that translated connectomes can substitute for real ones in SVM classification on some tasks, though not all.

Load-bearing premise

The method assumes the 116 brain regions in the AAL atlas are ordered so that neighboring entries in the connectivity matrix are anatomically related; if that ordering is arbitrary, the convolutional layers are just learning a fixed permutation and the spatial-structure advantage disappears.

Editorial extensions

If this is right

  • Subjects with only dMRI or only fMRI can have the missing connectome generated, making joint SC-FC analyses possible on cohorts where one modality was not acquired.
  • The direction-specific generators provide individual-level structure-function maps, since each subject's translated connectome retains its own topology rather than collapsing to an average brain.
  • If translated connectomes preserve graph metrics, they can be used in downstream network analyses such as density, efficiency, or modularity comparisons without re-acquiring the missing scan.
  • Adding the structure-preserving loss is a direct improvement over the same architecture without it, so the loss itself is a reusable component for other connectome-generation models.

Reading between the lines

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

  • A permutation test on the atlas ordering would separate the convolutional spatial-structure effect from the network merely learning a fixed grid permutation; the paper does not report this.
  • Because the architecture and losses are atlas-agnostic, the same bidirectional translation approach could be applied at other parcellations or to connectivity matrices from other modalities; the symmetry-preserving output layer would carry over unchanged.
  • The classification results vary by dataset and translation direction, so the practical promise of imputation should be validated per application; a translated SC that works for ADNI may not automatically work for DUMC-MDD.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes SFC-GAN, a CycleGAN-based framework for bidirectional translation between structural connectivity (SC) and functional connectivity (FC) matrices derived from the AAL atlas (116 regions). The method uses convolutional generators and discriminators, a cycle-consistency loss, an identity loss, and a structure-preserving (MSE + Pearson correlation) loss. The authors evaluate translated connectomes on ADNI (120 subjects) and DUMC-MDD (43 subjects) using similarity metrics (MSE, MAE, SSIM, Pearson, cosine) and graph-property absolute percentage differences, and report SVM classification accuracy on real versus translated connectomes. The central claim is that SFC-GAN outperforms baseline models in similarity and graph-property evaluations and that the translated connectomes can be effectively used for downstream classification.

Significance. If substantiated, a reliable bidirectional SC-FC translator would be practically valuable because many clinical datasets contain only one modality, and imputing the missing connectome would enable joint structure-function analyses. The paper also addresses a real gap: most prior work performs unidirectional prediction or single-modality augmentation. The architecture is a reasonable adaptation of CycleGAN, with explicit attention to symmetry preservation via a symmetric-output layer, and the structure-preserving loss is a sensible inductive bias. However, the manuscript's empirical validation is currently insufficient to support the claimed superiority, and several reported numbers contradict the paper's own conclusions. The absence of any baseline comparison is load-bearing, and the classification results in Table II do not support the 'effectively utilized' claim. The paper's contribution is therefore not yet established.

major comments (4)
  1. [Section III-B, Tables I-II] The central claim of 'superior performance... outperforming baseline models' is unsupported because no baseline models are reported. Table I compares only SFC-GAN with and without the proposed LSP loss, which is an ablation rather than a comparison against existing SC/FC prediction methods (e.g., Zhang et al. [15], Multi-GCN [16], BrainNetGAN [9]) or even a trivial mean-imputation baseline. Without such baselines, absolute similarity values and graph-property APDs have no reference point for establishing superiority. This is a load-bearing omission that directly invalidates the abstract's claim.
  2. [Eq. (2), Section II-B] The cycle-consistency loss is mis-specified: both terms use tilde_x_SC, so the second term should be Ex_FC [||G_FC(G_SC(x_FC)) - x_FC||_1] or equivalently tilde_x_FC, not tilde_x_SC. As written, the loss does not enforce cycle consistency for the FC domain. This appears to be a typo, but it obscures the actual objective and should be corrected.
  3. [Table II, Section III-B] The classification results contradict the claim that translated connectomes are 'well-suited for subsequent analyses.' On ADNI, translated FC accuracy is 16.67% versus 66.67% for real FC; on DUMC-MDD, translated SC accuracy is 23.08% versus 53.85% for real SC. These are severe degradations, and the paper's own text acknowledges these 'inferior' results but then concludes that translated connectomes are well-suited. The conclusion is not supported by the reported numbers, and the pattern is inconsistent (e.g., combined translated FC+SC on ADNI shows higher AUC than real FC+SC, which is difficult to interpret).
  4. [Table I, Section III-B] The statement that adding LSP 'enhances performance across all metrics' is contradicted by the reported numbers. For DUMC-MDD translated FC, MSE worsens from 0.0210 to 0.0237 and MAE from 0.1141 to 0.1224 when LSP is added; for ADNI translated SC, Pearson correlation and cosine similarity also do not improve (e.g., Pearson drops from 4.03 to 55.63? Actually the table shows a large jump, but the direction of improvement is not consistent). The claim of uniform improvement is therefore inaccurate and needs to be qualified or corrected.
minor comments (4)
  1. [Section II-B, Fig. 1c] The use of 3x3 and 4x4 convolutional kernels on a 116x116 matrix presupposes that the AAL atlas node ordering places spatially or functionally related regions adjacently. This assumption is not stated or validated; if the ordering is arbitrary, the claimed spatial-structure advantage of convolutions is not well founded. A discussion or ablation with a shuffled ordering would clarify the mechanism.
  2. [Section III-A] The dataset sizes are very small (96/24 split for ADNI, 30/13 for DUMC-MDD), and no statistical significance testing is reported for the similarity or graph-property differences. The standard deviations are large relative to some reported differences, and the reader cannot assess whether improvements are reliable.
  3. [Table I] The table would benefit from clear column units and a note on whether SSIM, Pearson, and cosine values are percentages or raw coefficients; the current formatting makes some entries hard to interpret (e.g., '4.03±1.85' for Pearson in Translated SC without LSP).
  4. [Throughout] There are several grammatical and typographical issues, including missing punctuation in the abstract ('data, each translated modality...'), inconsistent notation for x_FC/x_fc, and undefined abbreviations such as 'LSP' in equations before its introduction. A careful proofreading pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SFC-GAN is a self-contained CycleGAN extension; its losses and evaluation metrics overlap, but the test targets are held out and not used to fit the generators.

full rationale

The derivation chain is empirical rather than definitional. The generators are trained on a train split (ADNI 96/24; DUMC-MDD 30/13) and the translated connectomes are compared to held-out ground-truth connectomes in Table I. The fact that LSP uses MSE and PCC (Eq. 4) and Table I evaluates MSE, MAE, SSIM, Pearson and cosine similarity is a metric-alignment choice; the evaluation quantities are not parameters fitted to the test targets, so the 'prediction' is not forced by construction. The self-citation to [12] for the symmetric generator architecture is a design provenance citation, not an evidence loop: the claimed translation quality is measured against ground truth, not derived from [12]. The abstract's 'outperforming baseline models' is unsubstantiated because no baselines appear in Tables I-II, and the classification results in Table II partly contradict the usefulness claim (e.g., ADNI translated FC 16.67% vs real FC 66.67%); however, missing baselines and internal inconsistencies are correctness/evidence flaws, not circularity. The AAL node-ordering concern about convolutional kernels is a validity assumption, not a self-referential reduction. No equation in the paper reduces to its own input, and no load-bearing claim is imported solely from a same-author citation. Therefore the circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The model relies on hand-set architecture capacities and equal loss weights, on the domain assumption of a learnable SC-FC mapping, and on the assumption that a fixed AAL parcellation makes convolutional filtering meaningful. No new physical entities are introduced.

free parameters (2)
  • Equal loss weighting = 1
    All five loss terms (adversarial, cycle, identity, structure-preserving) are summed with equal weight in Eq. (7); this weighting is hand-set and not ablated.
  • Architecture capacities (filter counts 64, 128, 256, 512) = various
    Chosen by hand; no ablation or sensitivity analysis is reported.
assumptions (3)
  • domain assumption FC and SC are strongly correlated and a deterministic subject-level mapping exists between them.
    Invoked in the Introduction citing [4] and is the premise for learnable translation; if the mapping is many-to-many or weak, the bidirectional translation task is ill-posed.
  • domain assumption AAL atlas parcellation into 116 ROIs provides a common, ordered node space for both modalities.
    Both FC and SC are represented as 116x116 matrices in the same AAL ordering; the validity of convolutional translation depends on this ordering being meaningful.
  • ad hoc to paper CycleGAN cycle-consistency and identity losses are appropriate for connectome translation.
    The paper adopts CycleGAN [20] without justifying that the cycle-consistency assumption holds for SC-FC pairs; no ablation of loss components beyond the single LSP on/off comparison is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SFC-GAN: A Generative Adversarial Network for Brain Functional and Structural Connectome Translation." pith.science (2026). https://pith.science/paper/KTIALLJA

@misc{pith2026250107055,
  author       = {Pith},
  title        = {Pith review of: SFC-GAN: A Generative Adversarial Network for Brain Functional and Structural Connectome Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KTIALLJA}},
  note         = {Machine review of arXiv:2501.07055}
}
read the original abstract

Modern brain imaging technologies have enabled the detailed reconstruction of human brain connectomes, capturing structural connectivity (SC) from diffusion MRI and functional connectivity (FC) from functional MRI. Understanding the intricate relationships between SC and FC is vital for gaining deeper insights into the brain's functional and organizational mechanisms. However, obtaining both SC and FC modalities simultaneously remains challenging, hindering comprehensive analyses. Existing deep generative models typically focus on synthesizing a single modality or unidirectional translation between FC and SC, thereby missing the potential benefits of bi-directional translation, especially in scenarios where only one connectome is available. Therefore, we propose Structural-Functional Connectivity GAN (SFC-GAN), a novel framework for bidirectional translation between SC and FC. This approach leverages the CycleGAN architecture, incorporating convolutional layers to effectively capture the spatial structures of brain connectomes. To preserve the topological integrity of these connectomes, we employ a structure-preserving loss that guides the model in capturing both global and local connectome patterns while maintaining symmetry. Our framework demonstrates superior performance in translating between SC and FC, outperforming baseline models in similarity and graph property evaluations compared to ground truth data, each translated modality can be effectively utilized for downstream classification.

Figures

Figures reproduced from arXiv: 2501.07055 by the authors.

Figure 1
Figure 1. Overview of SFC-GAN: a Training of SFC-GAN, GSC translates FC, xf c to SC x˜sc, where GF C translates SC, xsc to FC x˜f c. DF C and DSC aim to discriminate between xf c and x˜f c, xsc and x˜sc, respectively. b Cycle consistency loss of FC and SC domains. c Network architectures of GF C , GSC , DF C , and DSC . loss inspired by [15], which guides the model in capturing both global and local connectome patterns. Addit… view at source ↗
Figure 2
Figure 2. Matrix reconstruction results on both ADNI and DUMC-MDD datasets using [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Reconstruction results on both ADNI and DUMC-MDD datasets with top 5% strongest connectivity in the brain space. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 26 canonical work pages

  1. [15]

    Predicting brain structural network using functional connectivity,

    L. Zhang, et al., “Predicting brain structural network using functional connectivity,” Medical image analysis , vol. 79, pp. 102463, 2022

    work page 2022

  2. [16]

    Recovering brain structural con- nectivity from functional connectivity via multi-gcn based generative adversarial network,

    L. Zhang, L. Wang, and D. Zhu, “Recovering brain structural con- nectivity from functional connectivity via multi-gcn based generative adversarial network,” in Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part VII 23 . Springer, 2020, pp. 53–61

    work page 2020

  3. [9]

    BrainNetGAN: Data augmentation of brain connectivity using generative adversarial network for dementia classification,

    C. Li, et al., “BrainNetGAN: Data augmentation of brain connectivity using generative adversarial network for dementia classification,” in Deep Generative Models, and Data Augmentation, Labelling, and Im- perfections. Springer, 2021, pp. 103–111

    work page 2021

  4. [1]

    Function-specific and enhanced brain structural connectivity mapping via joint modeling of diffusion and functional mri,

    S.-H. Chu, K. K. Parhi, and C. Lenglet, “Function-specific and enhanced brain structural connectivity mapping via joint modeling of diffusion and functional mri,” Scientific reports, vol. 8, no. 1, pp. 4741, 2018

    work page 2018

  5. [2]

    Brain connectivity studies on structure-function relationships: A short survey with an emphasis on machine learning,

    S. Wein, et al., “Brain connectivity studies on structure-function relationships: A short survey with an emphasis on machine learning,” Computational intelligence and neuroscience , vol. 2021, no. 1, pp. 5573740, 2021

    work page 2021

  6. [3]

    Local structure-function relationships in human brain networks across the lifespan,

    F. Zamani Esfahlani, et al., “Local structure-function relationships in human brain networks across the lifespan,” Nature communications , vol. 13, no. 1, pp. 2053, 2022

    work page 2022

  7. [4]

    Measuring brain connectivity: diffusion tensor imaging validates resting state temporal correlations,

    P. Skudlarski, et al., “Measuring brain connectivity: diffusion tensor imaging validates resting state temporal correlations,” Neuroimage, vol. 43, no. 3, pp. 554–561, 2008

    work page 2008

  8. [5]

    Identification of mci individuals using structural and functional connectivity networks,

    C.-Y . Wee, et al., “Identification of mci individuals using structural and functional connectivity networks,” Neuroimage, vol. 59, no. 3, pp. 2045–2056, 2012

    work page 2012

Show all 28 references
  1. [6]

    Machine learning based on functional and structural con- nectivity in mild cognitive impairment,

    Y . Li, et al., “Machine learning based on functional and structural con- nectivity in mild cognitive impairment,” Magnetic Resonance Imaging , vol. 109, pp. 10–17, 2024

    2024

  2. [7]

    A unified framework for static and dynamic functional connectivity augmentation for multi-domain brain disorder classifica- tion,

    Y .-F. Tan, et al., “A unified framework for static and dynamic functional connectivity augmentation for multi-domain brain disorder classifica- tion,” in 2023 IEEE International Conference on Image Processing (ICIP). IEEE, 2023, pp. 635–639

    2023

  3. [8]

    Alzheimer’s disease prediction via brain structural- functional deep fusing network,

    Q. Zuo, et al., “Alzheimer’s disease prediction via brain structural- functional deep fusing network,” IEEE Transactions on Neural Systems and Rehabilitation Engineering , vol. 31, pp. 4601–4612, 2023

    2023

  4. [10]

    Brain functional connectivity augmentation method for mental disease classification with generative adversarial network,

    Q. Yao and H. Lu, “Brain functional connectivity augmentation method for mental disease classification with generative adversarial network,” in PRCV. Springer, 2019, pp. 444–455

    2019

  5. [11]

    Data augmentation using generative adversarial neural networks on brain structural connectivity in multiple sclerosis,

    B. Barile, et al., “Data augmentation using generative adversarial neural networks on brain structural connectivity in multiple sclerosis,” Comput. Methods and Programs in Biomedicine , vol. 206, pp. 106113, 2021

    2021

  6. [12]

    fMRI Functional Connectivity Augmentation Using Convolutional Generative Adversarial Networks for Brain Disorder Classification,

    Y .-F. Tan, et al., “fMRI Functional Connectivity Augmentation Using Convolutional Generative Adversarial Networks for Brain Disorder Classification,” in 2024 IEEE International Symposium on Biomedical Imaging (ISBI). IEEE, 2024, pp. 1–5

    2024

  7. [13]

    Brainfc-cgan: A conditional generative adversar- ial network for brain functional connectivity augmentation and aging synthesis,

    Y .-F. Tan, et al., “Brainfc-cgan: A conditional generative adversar- ial network for brain functional connectivity augmentation and aging synthesis,” in ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 1511–1515

    2024

  8. [14]

    Adversarial learning based structural brain-network generative model for analyzing mild cognitive impairment,

    H. Kong, et al., “Adversarial learning based structural brain-network generative model for analyzing mild cognitive impairment,” in Chi- nese Conference on Pattern Recognition and Computer Vision (PRCV) . Springer, 2022, pp. 361–375

    2022

  9. [17]

    Structure can predict function in the human brain: a graph neural network deep learning model of functional connectivity and centrality based on structural connectivity,

    J. Neudorf, S. Kress, and R. Borowsky, “Structure can predict function in the human brain: a graph neural network deep learning model of functional connectivity and centrality based on structural connectivity,” Brain Structure and Function , vol. 227, no. 1, pp. 331–343, 2022

    2022

  10. [18]

    Learning to model the relationship between brain structural and functional connectomes,

    Y . Li, G. Mateos, and Z. Zhang, “Learning to model the relationship between brain structural and functional connectomes,” IEEE Transac- tions on Signal and Information Processing over Networks , vol. 8, pp. 830–843, 2022

    2022

  11. [19]

    Supervised graph representation learning for modeling the relationship between structural and functional brain connectivity,

    Y . Li, et al., “Supervised graph representation learning for modeling the relationship between structural and functional brain connectivity,” in ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2020, pp. 9065–9069

    2020

  12. [20]

    Unpaired image-to-image translation using cycle- consistent adversarial networks,

    J.-Y . Zhu, et al., “Unpaired image-to-image translation using cycle- consistent adversarial networks,” in Proceedings of the IEEE interna- tional conference on computer vision , 2017, pp. 2223–2232

    2017

  13. [21]

    Reduced default mode network functional connec- tivity in patients with recurrent major depressive disorder,

    C.-G. Yan, et al., “Reduced default mode network functional connec- tivity in patients with recurrent major depressive disorder,” Proc. of the National Academy of Sci. , vol. 116, no. 18, pp. 9078–9083, 2019

    2019

  14. [22]

    DPARSF: a matlab toolbox for

    C. Yan and Y . Zang, “DPARSF: a matlab toolbox for” pipeline” data analysis of resting-state fMRI,” Frontiers in Syst. Neurosci., p. 13, 2010

    2010

  15. [23]

    Panda: a pipeline toolbox for analyzing brain diffusion images,

    Z. Cui, et al., “Panda: a pipeline toolbox for analyzing brain diffusion images,” Frontiers in human neuroscience , vol. 7, pp. 42, 2013

    2013

  16. [24]

    Brain network functional connectivity and cognitive performance in major depressive disorder,

    K. M. Albert, et al., “Brain network functional connectivity and cognitive performance in major depressive disorder,” J. Psychiatr. Res., vol. 110, pp. 51–56, 2019

    2019

  17. [25]

    A bayesian approach to examining default mode network functional connectivity and cognitive performance in major depressive disorder,

    R. Wang, et al., “A bayesian approach to examining default mode network functional connectivity and cognitive performance in major depressive disorder,” Psychiatry Res. Neuroimaging , vol. 301, pp. 111102, 2020

    2020

  18. [26]

    LIBSVM: A library for support vector machines,

    C.-C. Chang and C.-J. Lin, “LIBSVM: A library for support vector machines,” ACM Trans. Intell. Syst. Technol. , vol. 2, no. 3, pp. 1–27, 2011

    2011

  19. [27]

    Tensorflow: Large-scale machine learning on hetero- geneous distributed systems,

    M. Abadi, et al., “Tensorflow: Large-scale machine learning on hetero- geneous distributed systems,” arXiv preprint arXiv:1603.04467 , 2016

    2016 arXiv

  20. [28]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 , 2014

    2014 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.