Structural Connectome Harmonization Using Deep Learning: The Strength of Graph Neural Networks

Anjali Tarun; Jagruti Patel; Jonas Richiardi; Mikkel Sch\"ottner; Patric Hagmann; Sebastien Tourbier; Thomas A. W. Bolton; Yasser Alem\`an-G\`omez

REVIEW 3 major objections 5 minor 6 references

Structural Connectome Harmonization Using Deep Learning: The Strength of Graph Neural Networks

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A graph autoencoder harmonizes brain connectomes across scanners using only the site label, and preserves each person's network fingerprint better than non-graph autoencoders.

desk verdict A careful simulation benchmark for graph-based connectome harmonization, but the domain-generalization claim runs ahead of the evidence. read the letter →

arxiv 2507.13992 v1 pith:3YLSPXV6 submitted 2025-07-18 cs.LG

Jagruti Patel , Thomas A. W. Bolton , Mikkel Schöttner , Anjali Tarun , Sebastien Tourbier , Yasser Alemàn-Gòmez , Jonas Richiardi , Patric Hagmann This is my paper · ORCID

classification cs.LG

keywords structuralconnectomeharmonizationgraphneuralnetworksautoencodermulti-sitestudiesdiffusionMRIbrainnetworkfingerprintingsite-invariantrepresentationtopology

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 argues that scanner-related bias in structural connectomes can be removed at the matrix level by a deep network that knows only which site produced each scan, without needing detailed acquisition metadata or traveling subjects. The core comparison is architectural: a graph autoencoder that treats the connectome as a network, a convolutional autoencoder that treats it as an image, and a fully connected autoencoder that treats it as a flat vector, all inside the same site-conditioned, adversarially trained framework. The graph version preserves graph-theoretic properties and subject-level identifiability better than the other deep models, while the linear-regression baseline still wins on raw numerical accuracy because it is given the acquisition parameters explicitly. The practical point is that when such metadata is unavailable, graph-based deep harmonization offers a structure-aware, domain-generalizable alternative for pooling multi-site data.

What carries the argument

The load-bearing mechanism is a site-conditioned graph autoencoder whose encoder stacks Chebyshev spectral graph-convolution layers (ChebConv), so each node's representation is updated by filtering on the normalized graph Laplacian with a localized polynomial filter. A gradient-reversal site classifier forces the encoder's node embeddings to be site-invariant, while a site-mapper encodes the target site label and an AdaIN (adaptive instance normalization) fusion module re-inserts site-specific statistics at each decoding stage. The decoder then reconstructs a harmonized structural connectivity matrix for the target site, blending the subject's invariant traits with target-site attributes.

What would settle it

Use traveling subjects scanned on genuinely different scanners and vendors, harmonize with the site-conditioned graph autoencoder trained on site labels only, and compare the harmonized connectomes to the same subjects' scans on a reference scanner; if fingerprinting accuracy and graph-metric error are no better than the unharmonized lower bound, or if site-specific clusters persist in the latent space, the central claim fails.

Watch

Extended reading notes

Core claim

The central discovery claimed is that model architecture determines what harmonization preserves. Trained to encode a site-invariant latent representation and decode it under a target-site condition, the graph autoencoder reconstructs low-weight, individually distinctive connections that image- and vector-based autoencoders wash out, and its latent space mixes sites while keeping the same subject's scans close together. It achieves near-perfect fingerprinting accuracy and the best identifiability difference among the deep models, while the convolutional autoencoder leads on edge-existence detection and the fully connected autoencoder on raw edge-weight error. The linear-regression baseline attains the best overall numbers by explicitly modeling resolution and b-value, but the paper concludes that its dependence on acquisition metadata makes it impractical for real multi-site studies, leaving the graph autoencoder as the recommended structure-aware harmonizer.

Load-bearing premise

The evaluation rests on the assumption that resampling one dataset's diffusion images to two spatial resolutions and two diffusion-weighting settings mimics the acquisition differences of real multi-site studies; if real scanner variation includes factors this simulation omits, the site-conditioned model's advantages may not transfer.

Editorial extensions

If this is right

Harmonization needs only the site label at inference, so multi-site pooling can proceed when acquisition metadata and traveling subjects are unavailable.
Graph topology is not a side detail: flattening the connectome (FAE) collapses fingerprinting accuracy to near zero, while graph-aware decoding keeps it near the test-retest upper bound.
Downstream task should decide the architecture: CAE for edge-existence and demographic prediction, GAE for fingerprinting and graph-metric preservation.
The single-encoder, single-decoder, site-conditioned design scales to additional sites without adding site-specific encoders, since site identity enters only through the conditioning vector.
Even the best deep models stay below linear regression on raw edge-weight error, so metadata-free harmonization currently trades some numerical fidelity for applicability.

Reading between the lines

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

A testable extension the authors leave implicit: apply the same site-conditioned GAE to genuine multi-site data with traveling-subject ground truth; if real scanner variation includes vendor, field strength, or sequence effects beyond resolution and b-value, the site label alone may or may not absorb them, and the GAE's edge-existence weakness could surface as spurious connections.
The GAE's preservation of low-weight connections suggests that group-level or case-control studies of subtle connectivity differences might gain sensitivity from graph-based harmonization, a claim the paper does not directly test.
One could combine the GAE's topological strengths with an explicit non-connection penalty or a weighting scheme to close the MAE gap toward the linear baseline without reintroducing metadata dependence.
The near-zero fingerprinting of the fully connected autoencoder isolates the cause: treating a connectome as a flat vector discards exactly the relational structure that makes a brain network identifiable, which argues against vectorizing connectomes in other harmonization pipelines.

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.

Desk Editor's Note

read the letter

Patel et al. test three deep autoencoder architectures for structural connectome harmonization—fully connected, convolutional, and graph—against a linear regression baseline in a simulated four-site design using HCP data. The new contribution is a graph convolutional autoencoder with AdaIN-based site conditioning and an adversarial site classifier, plus a careful head-to-head evaluation. The execution is honest: family-based data splitting, test-retest subjects kept out of training, held-out ground truth, and metrics spanning edge weights, topology, fingerprinting, and demographics.

What it does well: the simulation is clearly documented, the augmentation check is sensible, and the main empirical finding—architecture matters, with the GAE preserving topology and subject identity better than the FAE or CAE while the LR baseline wins when full acquisition metadata is available—is useful and plausible. The t-SNE and qualitative examples support the disentanglement story.

The soft spots are real but not fatal. The abstract and conclusion describe the framework as 'domain-generalizable,' yet the evidence consists of four sites that are deterministic functions of two acquisition parameters (resolution and b-value), with no held-out acquisition configuration, no real scanner variability, and only one transfer direction tested. Section 4.1 concedes that generalizability remains to be established, so the abstract overstates what the study shows. Second, the GAE advantage over CAE on several graph metrics is presented as point estimates without significance testing or interval overlap analysis; the confidence intervals in Table 2 suggest some differences are slim. Third, the GitHub link in the Data Availability statement is malformed, which effectively makes the code inaccessible.

If read as a controlled simulation benchmark, the central argument holds up. The paper does not demonstrate domain-generalizable harmonization, but it does make a credible case that graph-based architectures are worth pursuing for structure-aware harmonization. I would not cite it as evidence of real-world generalizability, but I would cite it as a simulation-based benchmark of deep harmonization architectures. A serious editor should send it to peer review; the necessary revisions are to soften the generalization claim (or add a held-out site) and to add statistical comparison across methods.

Referee Report

3 major / 5 minor

Summary. The paper proposes a unified site-conditioned deep harmonization framework for structural connectomes, instantiated as a fully connected autoencoder (FAE), a 2D convolutional autoencoder (CAE), and a graph convolutional autoencoder (GAE), and compares them against a linear regression (LR) baseline. The models are trained adversarially with a site classifier and a site-conditioned decoder, using only site identity at inference. Evaluation is performed on a simulated four-site scenario derived from HCP-YA data by resampling DWI to 2.3 mm and extracting b-value shells at 1000 and 3000, with harmonization from the lowest-quality acquisition to the highest-quality acquisition. The paper reports that LR achieves the best numerical performance on most metrics, while GAE best preserves graph topology and subject-level individuality, as measured by graph metrics and fingerprinting/identifiability.

Significance. If the central claims are established, the framework would be practically valuable for multi-site neuroimaging studies because it promises harmonization without detailed acquisition metadata or traveling subjects. The study has notable strengths: the data split avoids family-level leakage, test-retest subjects are excluded from training, evaluation is performed against held-out ground-truth connectomes, multiple complementary metrics are used, lower and upper performance bounds are defined, and code is made available. However, the significance is currently bounded by three issues: the absence of statistical testing for the headline architecture comparisons, the evaluation of only one source-to-target transfer direction, and the simulated nature of the four sites, which the paper's own Limitations section notes does not establish real-world generalizability.

major comments (3)

[§3.1.2, Table 2] The central claim that GAE 'consistently outperforms' CAE on topological preservation and individuality rests on point estimates without significance testing. The differences are small relative to the per-subject standard deviations: NS 641.99±177.98 vs 652.29±200.30, CLC 3.81±0.36 vs 4.05±0.71, LE 6.36±0.56 vs 8.40±1.49, and FA/ID are reported with no dispersion or confidence interval at all. A fingerprinting difference of 0.99 vs 0.96 corresponds to roughly three subjects out of 107 and could easily arise by chance. Since the abstract's superiority claim is a primary contribution, the authors should add paired statistical tests across test subjects (e.g., bootstrap or permutation tests on per-subject metric differences) and report effect sizes or confidence intervals for all headline metrics.
[§2.3, Table 2; §1 Contributions] The evaluation covers exactly one transfer direction: the lowest-quality site (b-value 1000, 2.3 mm, Site 0) is harmonized to the highest-quality site (b-value 3000, 1.25 mm, Site 3). Contribution 1 claims the framework accurately predicts SCs 'from any source site to any target site,' and Section 2.2.3 describes inference with arbitrary target site vectors, but no results are shown for other source sites (Sites 1 or 2), for other target sites, or for a leave-one-site-out protocol. The cross-site generality claim is therefore supported only for a single low-to-high quality direction. Please either report the full source-target matrix of harmonization results or explicitly restrict the claims to this directional scenario.
[Abstract and §4.1] The abstract and conclusion describe the method as 'domain-generalizable,' but the protocol contains no held-out acquisition configuration or unseen site. The four sites are generated by resampling the same HCP-YA diffusion data along two axes (spatial resolution and b-value shell), so the site label is a deterministic one-hot encoding of exactly those two acquisition parameters. The Limitations section itself states that 'the generalizability of these findings ... remains to be fully established.' This is a load-bearing gap because the practical value of the framework lies in its applicability to real multi-site data with vendor, field-strength, sequence, and noise differences. I recommend either softening the generalization claims throughout, or adding an experiment that holds out an entire acquisition configuration and evaluates harmonization to it, ideally on data with more than two manipulated acquisition parameters.

minor comments (5)

[Abstract] The sentence 'Large-scale multi-site studies have exist' contains a grammatical error and should read 'have existed' or simply 'exist.'
[§2.2.3, CAE and GAE paragraphs] Both the CAE and GAE subsections end with 'Detailed architecture of FAE is provided in Supplementary Figure 2/3'; these should refer to the CAE and GAE architectures, respectively.
[§1 Contributions] The phrase 'The framework adapts incorporates data augmentation techniques' has a duplicated verb; it should be 'adapts' or 'incorporates,' not both.
[§2.1.3] The target site is described only as 'highest b-value of 3000 and highest isotropic spatial resolution of 1.25 mm'; identifying this explicitly as Site 3 in Table 1 would improve clarity, given that the evaluation uses a single fixed target.
[Table 1] The header 'Spatial resolution (mm3)' is confusing; since resolution is isotropic, the unit should be 'mm' or 'mm^3' only if a volume is meant, and the latter does not match the text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the GAE harmonization outputs are evaluated against held-out ground-truth connectomes, and the site-conditioned vectors are inputs to the model, not fitted predictions.

full rationale

The paper's derivation chain is self-contained and empirically testable rather than circular. The four sites are simulated by resampling the same HCP data (Section 2.1), and Table 1 defines the site-condition variable as a combination of b-value and resolution, but this condition vector is an input used to drive the decoder during inference, not a quantity derived from the evaluation target. Section 2.2.3 states that during validation and testing, matrices are harmonized 'by adjusting vc to the target site with the highest quality,' and the resulting outputs are compared against the held-out high-quality S_high matrices. No equation sets the reported harmonization metric equal to a fitted parameter or reduces the predicted connectome to the conditioning vector. The LR baseline from Patel et al. (2024) is explicitly formulated and benchmarked as a comparison method, and the self-citations to Patel et al. provide preprocessing provenance or explain augmentation behavior rather than supplying a load-bearing justification for the GAE's central claim. The paper's own limitation statement in Section 4.1—that 'the generalizability of these findings... remains to be fully established'—is an external-validity caveat about the simulated design, not evidence that the reported harmonization results are defined into existence. Overall, the central empirical comparison (GAE versus FAE, CAE, and LR on topological and fingerprinting metrics) is not circular; it is a genuine evaluation against held-out ground truth, and any concerns about real-world generalization belong to correctness risk rather than circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The core contribution is an empirical framework, so the ledger contains learned site-conditioning vectors and hand-chosen training constants, plus domain assumptions about the simulation fidelity and the ground-truth target. No new physical or mathematical entities are introduced.

free parameters (3)

Site-conditioning latent vectors fM = 4 vectors, dimension K (not reported)
Each site index is mapped to a learned latent vector that modulates the decoder via AdaIN or concatenation; these vectors encode site-specific acquisition effects and directly determine harmonization fidelity.
MAE loss weighting for existing edges = 2.5
The reconstruction loss weights existing edge-weight errors 2.5x relative to non-connections; this hand-chosen constant shapes the edge-weight/edge-existence trade-off across models.
Augmentation count per site = 4000
The number of Mixup-style augmented matrices generated per acquisition parameter combination; this hand-chosen data-engineering quantity affects the training distribution.

assumptions (5)

domain assumption The HCP DWI resampling protocol (downsampling to 2.3mm and selecting b-value shells 1000 and 3000) reproduces the acquisition-related variation observed across real multi-site studies.
Section 2.1 states the resampling 'effectively mimicked a scenario where each subject was scanned under four distinct acquisition parameters'; the entire evaluation rests on this premise.
domain assumption The b-value 3000 and 1.25mm acquisition is the correct high-quality ground truth target for harmonization.
Section 2.1.3 defines the highest-quality SCs as the harmonization target; if this acquisition is not biologically or practically superior, the evaluation target is arbitrary.
standard math Deterministic tractography with 10 million streamlines and the Lausanne 2018 scale-3 atlas provides a valid structural connectome representation.
These standard pipelines are invoked in Section 2.1 without fresh justification.
standard math Chebyshev polynomial graph convolutions and AdaIN correctly implement the site-conditioned generative mapping.
Section 2.2.3 adopts these standard building blocks without proof; they are assumed to behave as described.
domain assumption Mixup-style augmentation preserves the distribution of subject-level connectome properties needed for harmonization.
Section 2.1.2 validates key graph metrics on augmented data but clustering coefficient is reduced, yet the model is still trained on the augmented distribution.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Structural Connectome Harmonization Using Deep Learning: The Strength of Graph Neural Networks." pith.science (2026). https://pith.science/paper/3YLSPXV6

@misc{pith2026250713992,
  author       = {Pith},
  title        = {Pith review of: Structural Connectome Harmonization Using Deep Learning: The Strength of Graph Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3YLSPXV6}},
  note         = {Machine review of arXiv:2507.13992}
}

read the original abstract

Small sample sizes in neuroimaging in general, and in structural connectome (SC) studies in particular limit the development of reliable biomarkers for neurological and psychiatric disorders - such as Alzheimer's disease and schizophrenia - by reducing statistical power, reliability, and generalizability. Large-scale multi-site studies have exist, but they have acquisition-related biases due to scanner heterogeneity, compromising imaging consistency and downstream analyses. While existing SC harmonization methods - such as linear regression (LR), ComBat, and deep learning techniques - mitigate these biases, they often rely on detailed metadata, traveling subjects (TS), or overlook the graph-topology of SCs. To address these limitations, we propose a site-conditioned deep harmonization framework that harmonizes SCs across diverse acquisition sites without requiring metadata or TS that we test in a simulated scenario based on the Human Connectome Dataset. Within this framework, we benchmark three deep architectures - a fully connected autoencoder (AE), a convolutional AE, and a graph convolutional AE - against a top-performing LR baseline. While non-graph models excel in edge-weight prediction and edge existence detection, the graph AE demonstrates superior preservation of topological structure and subject-level individuality, as reflected by graph metrics and fingerprinting accuracy, respectively. Although the LR baseline achieves the highest numerical performance by explicitly modeling acquisition parameters, it lacks applicability to real-world multi-site use cases as detailed acquisition metadata is often unavailable. Our results highlight the critical role of model architecture in SC harmonization performance and demonstrate that graph-based approaches are particularly well-suited for structure-aware, domain-generalizable SC harmonization in large-scale multi-site SC studies.

Figures

Figures reproduced from arXiv: 2507.13992 by the authors.

**Figure 2.** Comparison of different graph metrics between original ( [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Overview of our deep SC harmonization framework, featuring an encoder-decoder architec [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Spyder plot comparing deep SC harmonization methods. Metrics are normalized to a [0, 1] range, with MAE, BMAE, EV, NS, CC, CLC, and LE inverted for consistency, where higher values indicate better performance. 3.1 Quantitative evaluation A comparative summary of FAE, C…

**Figure 5.** Figure 5: Distribution of topological preservation metrics across test subjects for the highest-quality SCs, [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]

**Figure 6.** Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]

**Figure 7.** Figure 7: t-SNE visualization of encoder embeddings from the deep SC harmonization framework for [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

6 extracted references · 4 canonical work pages

[1]

Amico, E., & Go˜ ni, J. (2018). The quest for identifiability in human functional connectomes. Scientific reports, 8 (1),

work page 2018
[122]

Ganin, Y., & Lempitsky, V. (2015). Unsupervised domain adaptation by backpropagation. International conference on machine learning , 1180–1189. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., March, M., & Lempitsky, V. (2016). Domain-adversarial training of neural networks. Journal of machine learning research , 17 (59), ...

arXiv 2015
[396]

Roca, V., Kuchcinski, G., Pruvo, J.-P., Manouvriez, D., Lopes, R., et al. (2025). Iguane: A 3d generalizable cyclegan for multicenter harmonization of brain mr images. Medical Image Analysis , 99, 103388. Rubinov, M., & Sporns, O. (2010). Complex network measures of brain connectivity: Uses and interpre- tations. Neuroimage, 52 (3), 1059–1069. Shen, R. S....

work page 2025
[4248]

Tournier, J.-D., Calamante, F., & Connelly, A. (2007). Robust determination of the fibre orientation distribution in diffusion mri: Non-negativity constrained super-resolved spherical deconvolution. Neuroimage, 35 (4), 1459–1472. Tournier, J.-D., Calamante, F., & Connelly, A. (2012). Mrtrix: Diffusion tractography in crossing fiber regions. International ...

work page 2007
[4892]

Najdenovska, E., Alem´ an-G´ omez, Y., Battistella, G., Descoteaux, M., Hagmann, P., Jacquemont, S., Maeder, P., Thiran, J.-P., Fornari, E., & Bach Cuadra, M. (2018). In-vivo probabilistic atlas of human thalamic nuclei based on diffusion-weighted magnetic resonance imaging. Scientific Data, 5 (1), 1–11. Newlin, N. R., Kanakaraj, P., Li, T., Pechman, K., ...

work page 2018
[8254]

L., Kiros, J

Ba, J. L., Kiros, J. R., & Hinton, G. E. (2016). Layer normalization. arXiv preprint arXiv:1607.06450. Basaia, S., Agosta, F., Cividini, C., Trojsi, F., Riva, N., Spinelli, E. G., Moglia, C., Femiano, C., Castel- novo, V., Canu, E., et al. (2020). Structural and functional brain connectome in motor neuron diseases: A multicenter mri study. Neurology, 95 (...

arXiv 2016

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Amico, E., & Go˜ ni, J. (2018). The quest for identifiability in human functional connectomes. Scientific reports, 8 (1),

work page 2018

[2] [122]

Ganin, Y., & Lempitsky, V. (2015). Unsupervised domain adaptation by backpropagation. International conference on machine learning , 1180–1189. Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., March, M., & Lempitsky, V. (2016). Domain-adversarial training of neural networks. Journal of machine learning research , 17 (59), ...

arXiv 2015

[3] [396]

Roca, V., Kuchcinski, G., Pruvo, J.-P., Manouvriez, D., Lopes, R., et al. (2025). Iguane: A 3d generalizable cyclegan for multicenter harmonization of brain mr images. Medical Image Analysis , 99, 103388. Rubinov, M., & Sporns, O. (2010). Complex network measures of brain connectivity: Uses and interpre- tations. Neuroimage, 52 (3), 1059–1069. Shen, R. S....

work page 2025

[4] [4248]

Tournier, J.-D., Calamante, F., & Connelly, A. (2007). Robust determination of the fibre orientation distribution in diffusion mri: Non-negativity constrained super-resolved spherical deconvolution. Neuroimage, 35 (4), 1459–1472. Tournier, J.-D., Calamante, F., & Connelly, A. (2012). Mrtrix: Diffusion tractography in crossing fiber regions. International ...

work page 2007

[5] [4892]

Najdenovska, E., Alem´ an-G´ omez, Y., Battistella, G., Descoteaux, M., Hagmann, P., Jacquemont, S., Maeder, P., Thiran, J.-P., Fornari, E., & Bach Cuadra, M. (2018). In-vivo probabilistic atlas of human thalamic nuclei based on diffusion-weighted magnetic resonance imaging. Scientific Data, 5 (1), 1–11. Newlin, N. R., Kanakaraj, P., Li, T., Pechman, K., ...

work page 2018

[6] [8254]

L., Kiros, J

Ba, J. L., Kiros, J. R., & Hinton, G. E. (2016). Layer normalization. arXiv preprint arXiv:1607.06450. Basaia, S., Agosta, F., Cividini, C., Trojsi, F., Riva, N., Spinelli, E. G., Moglia, C., Femiano, C., Castel- novo, V., Canu, E., et al. (2020). Structural and functional brain connectome in motor neuron diseases: A multicenter mri study. Neurology, 95 (...

arXiv 2016