{"id":"abccb6c7-eaa1-40c9-a7d1-403eaf1f380c","arxiv_id":"2501.07055","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"SFC-GAN is a bidirectional CycleGAN for translating between structural and functional brain connectomes while aiming to preserve connectome topology.","lead":"The paper introduces SFC-GAN, a CycleGAN-based model that translates structural connectivity matrices into functional connectivity matrices and back. It targets situations where only one brain imaging modality is available, by generating the missing connectome for downstream analysis.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that SFC-GAN outperforms baseline models is unsubstantiated: no baseline models appear in any experiment, and internal classification results (ADNI translated FC 16.67% vs real FC 66.67%; DUMC translated SC 23.08% vs real SC 53.85%) directly contradict the usefulness claim.","rationale":"I read the paper as claiming an empirical advance: bidirectional SC/FC translation using a CycleGAN with convolutional generators and a structure-preserving loss, evaluated by similarity/graph metrics and downstream SVM classification. For that claim to hold, the reported experiments must demonstrate that the proposed model beats reasonable alternatives. They do not: Table I reports only the proposed model with and without LSP, and no comparison method appears anywhere in the paper. This is not a stylistic omission; it is the difference between 'competitive results' and the stated 'outperforming baseline models.' The classification table is not merely missing a comparison; it contains direct counterevidence. On ADNI, translated FC accuracy drops to 16.67% from 66.67% real-FC accuracy, and on DUMC-MDD translated SC accuracy drops to 23.08% from 53.85% real-SC accuracy. The paper's own text acknowledges 'inferior classification performance' for two settings but still concludes that translated connectomes are well-suited for downstream analyses, an internal tension in the central claim. A further internal inconsistency is the claim that LSP improves all metrics, while DUMC-MDD translated FC MSE and MAE worsen with LSP in Table I. I did not base the verdict on the reader's formal 'weakest assumption' about AAL node ordering and convolutional kernels. That concern targets the mechanism and is plausible, but it is not needed: even if the convolutional operators are accepted at face value, the empirical superiority claim has no baseline support. I agree with the reader's REJECT verdict, so no verdict change is needed. The concrete test would settle the issue by supplying the missing comparisons on the same splits.","tokens_in":7392,"tokens_out":4329,"duration_ms":41280,"concrete_test":"Run the identical train/test splits with three controls: (1) vanilla CycleGAN without LSP or identity loss; (2) a linear ridge regression mapping vectorized FC to SC and SC to FC; (3) an existing SC-FC prediction method such as Zhang et al. [15] or Multi-GCN [16] adapted to both directions. Compute MSE, MAE, SSIM, Pearson correlation, cosine similarity, and the four graph APDs on the same 24 ADNI and 13 DUMC-MDD test subjects, with paired bootstrap confidence intervals over 1000 resamples. Also report SVM classification accuracy and AUC for each baseline's translated connectomes. If SFC-GAN does not beat the best control on a majority of metrics across both datasets, the abstract's superiority claim and the 'effectively utilized for downstream classification' claim fail.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim is that SFC-GAN 'outperforms baseline models in similarity and graph property evaluations' and that translated connectomes can be effectively used for downstream classification. Section III-B and Table I compare only SFC-GAN with and without the proposed LSP loss; no CycleGAN variant without the proposed losses, no prior SC/FC prediction model (e.g., Zhang et al. [15], BrainNetGAN [9], Multi-GCN [16]), and no trivial imputation baseline is reported. Thus the superiority clause of the central claim is not testable from the paper. The classification evidence is internally problematic: in Table II, ADNI translated FC accuracy is 16.67% versus 66.67% for real FC, and DUMC-MDD translated SC accuracy is 23.08% versus 53.85% for real SC, while the paper nevertheless concludes that translated connectomes are well-suited for downstream analysis. Table I also contradicts the statement that LSP improves all metrics: DUMC-MDD translated FC MSE worsens from 0.0210 to 0.0237 and MAE from 0.1141 to 0.1224 when LSP is added. The absence of baselines is load-bearing because absolute similarity values and graph APDs have no reference point for establishing 'superior performance.' The reader's concern about AAL node ordering and convolutional kernels is relevant to the mechanism, but even granting that mechanism, the empirical claim remains unverified without baselines.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":7770,"tokens_out":2785,"duration_ms":27059,"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":[{"comment":"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.","section":"Section III-B, Tables I-II"},{"comment":"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.","section":"Eq. (2), Section II-B"},{"comment":"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).","section":"Table II, Section III-B"},{"comment":"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.","section":"Table I, Section III-B"}],"minor_comments":[{"comment":"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.","section":"Section II-B, Fig. 1c"},{"comment":"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.","section":"Section III-A"},{"comment":"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).","section":"Table I"},{"comment":"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.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The paper addresses an interesting and timely problem, and the SFC-GAN architecture is a plausible starting point. However, the empirical evidence is not sufficient for publication: the headline claim of outperforming baselines is untested, the classification results contradict the main conclusion, and there are internal inconsistencies in the reported metrics. These issues are load-bearing and would require new experiments (baseline comparisons, corrected evaluation) and a rewritten discussion. Given the current state, I cannot recommend acceptance or even minor revision. If the authors conduct a thorough comparison and fix the inconsistencies, a future submission could be worth considering."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nI read the SFC-GAN paper. The headline claim—bidirectional SC/FC translation that outperforms baselines—is not something you can verify from the paper, because there are no baselines. The experiments compare SFC-GAN with and without their structure-preserving loss (LSP), plus some classification numbers. Against CycleGAN or any prior SC/FC prediction model, nothing.\n\nWhat's new here is legitimate: it's the first attempt, per the citations, to do both directions (FC->SC and SC->FC) rather than one-way prediction. The architecture is a fairly straightforward CycleGAN adaptation with convolutional layers and a symmetry-preserving trick, and the LSP loss is borrowed from Zhang et al. but repurposed for translation. If you need missing-modality imputation, this is a sensible engineering direction.\n\nThe soft spots are not minor. Table I shows that adding LSP actually worsens MSE and MAE for DUMC-MDD translated FC (0.0210 to 0.0237 and 0.1141 to 0.1224), yet the text says LSP 'enhances performance across all metrics.' That's a direct contradiction. Table II shows translated FC on ADNI at 16.67% accuracy versus 66.67% for real FC, and translated SC on DUMC-MDD at 23.08% versus 53.85%, yet the conclusion says the translated connectomes are 'well-suited for downstream analysis.' The authors do acknowledge these failures in one sentence, but then walk it back. Eq. (2) has a typo that makes the cycle loss read as comparing the same variable twice. And no code or data is provided.\n\nThe reader flagged the AAL node ordering issue for convolutions. That's a real mechanistic concern, but I'd treat it as secondary: even if convolutions are just learning on a permuted grid, the empirical question is whether the translation works. The bigger problem is the evaluation doesn't give you a yardstick.\n\nBottom line: the paper would be useful to neuroimaging practitioners combining or imputing connectomes, but the central claim is unsupported and the internal consistency needs fixing. I'd give it a major-revision style review rather than desk reject, because the idea is sound and the fix is mostly about doing the evaluation properly. Just don't cite the numbers yet.","headline":"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.","tokens_in":8204,"tokens_out":2585,"would_cite":false,"duration_ms":24663,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["brain connectome","structural connectivity","functional connectivity","bidirectional translation","CycleGAN","structure-preserving loss","generative adversarial network","connectome topology"],"falsifier":"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.","tokens_in":7243,"feed_emoji":"🧠","tokens_out":7395,"duration_ms":68179,"temperature":0.7,"pith_summary":"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.","feed_headline":"Missing brain network maps are imputed from the other scan","feed_subtitle":"Bidirectional GAN translation keeps graph topology close to the ground truth.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Provides the CycleGAN backbone that SFC-GAN extends for bidirectional unpaired translation.","marker":"[20]"},{"why":"Source of the structure-preserving loss idea, combining MSE and Pearson correlation, and a unidirectional SC-from-FC baseline.","marker":"[15]"},{"why":"Supplies the symmetric generator design that SFC-GAN adopts to output symmetric connectomes.","marker":"[12]"},{"why":"Reference for processing ADNI resting-state fMRI data into functional connectomes.","marker":"[21]"},{"why":"Describes the fMRI preprocessing pipeline applied to ADNI data.","marker":"[22]"},{"why":"Pipeline used to process diffusion MRI into structural connectomes for the ADNI dataset.","marker":"[23]"},{"why":"Describes preprocessing and subject details for the DUMC-MDD dataset.","marker":"[24]"},{"why":"Additional reference for DUMC-MDD data preprocessing and quality control.","marker":"[25]"},{"why":"SVM classifier used to measure whether translated connectomes preserve diagnostic information.","marker":"[26]"}],"fun_headline_variants":["Bidirectional GAN generates missing brain connectomes","GAN keeps brain network topology intact in translation","CycleGAN translates between structural and functional scans","SFC-GAN imputes connectomes with structure-preserving loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Bidirectional GAN generates missing brain connectomes","GAN keeps brain network topology intact in translation","CycleGAN translates between structural and functional scans","SFC-GAN imputes connectomes with structure-preserving loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000356,"raw_usage":{"total_tokens":1925,"prompt_tokens":933,"completion_tokens":992,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":930}},"tokens_in":549,"tokens_out":992,"duration_ms":8375,"temperature":1.0,"reasoning_tokens":930,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:48:50.994227+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"A bayesian approach to examining default mode network functional connectivity and cognitive performance in major depressive disorder,","cited_arxiv_id":null,"evidence_quote":"Additional reference for DUMC-MDD data preprocessing and quality control."},{"cited_title":"Unpaired image-to-image translation using cycle- consistent adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Provides the CycleGAN backbone that SFC-GAN extends for bidirectional unpaired translation."},{"cited_title":"Predicting brain structural network using functional connectivity,","cited_arxiv_id":null,"evidence_quote":"Source of the structure-preserving loss idea, combining MSE and Pearson correlation, and a unidirectional SC-from-FC baseline."},{"cited_title":"fMRI Functional Connectivity Augmentation Using Convolutional Generative Adversarial Networks for Brain Disorder Classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the symmetric generator design that SFC-GAN adopts to output symmetric connectomes."},{"cited_title":"Reduced default mode network functional connec- tivity in patients with recurrent major depressive disorder,","cited_arxiv_id":null,"evidence_quote":"Reference for processing ADNI resting-state fMRI data into functional connectomes."},{"cited_title":"DPARSF: a matlab toolbox for","cited_arxiv_id":null,"evidence_quote":"Describes the fMRI preprocessing pipeline applied to ADNI data."},{"cited_title":"Panda: a pipeline toolbox for analyzing brain diffusion images,","cited_arxiv_id":null,"evidence_quote":"Pipeline used to process diffusion MRI into structural connectomes for the ADNI dataset."},{"cited_title":"Brain network functional connectivity and cognitive performance in major depressive disorder,","cited_arxiv_id":null,"evidence_quote":"Describes preprocessing and subject details for the DUMC-MDD dataset."},{"cited_title":"LIBSVM: A library for support vector machines,","cited_arxiv_id":null,"evidence_quote":"SVM classifier used to measure whether translated connectomes preserve diagnostic information."}],"review_version":1}