A Confounding Factors-Inhibition Adversarial Learning Framework for Multi-site fMRI Mental Disorder Identification

Bin Wan; Jie Xiang; Peng Wei; Rui Cao; Shijie Guo; Wenbo Ning; Xiaobo Liu; Xin Wen; Yan Niu

arxiv: 2504.09179 · v1 · submitted 2025-04-12 · 💻 cs.LG · cs.AI

A Confounding Factors-Inhibition Adversarial Learning Framework for Multi-site fMRI Mental Disorder Identification

Xin Wen , Shijie Guo , Wenbo Ning , Rui Cao , Yan Niu , Bin Wan , Peng Wei , Xiaobo Liu

show 1 more author

Jie Xiang

This is my paper

Pith reviewed 2026-05-22 20:56 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords multi-site fMRIadversarial learningfunctional connectivitymental disorder classificationsite variabilitynode information assemblyABIDE datasetADHD-200 dataset

0 comments

The pith

Adversarial network balances disorder classification against site regression to cut confounding effects in multi-site fMRI data.

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

The paper presents MSalNET, a framework that learns functional connectivity features from fMRI scans collected at multiple sites while actively suppressing site-specific confounding signals. It does so through a node information assembly step that gathers edge data along two axes to form node representations, a site feature extractor that operates directly on individual scans, and an adversarial setup governed by a custom loss that pits the main disorder classification task against a site regression task. A sympathetic reader would care because multi-site datasets normally suffer from scanner and population differences that degrade model reliability; removing those differences in a data-driven manner could let the same model work on new sites without manual harmonization. If the balance holds, classification should improve while site information is stripped from the learned features.

Core claim

The authors claim that jointly optimizing a disorder classifier and a site regressor inside an adversarial network, guided by a novel loss function, removes site variability from functional connectivity representations without erasing the signals needed for accurate mental disorder identification. This is realized by first assembling node features from horizontal and vertical edge information, then extracting site-level features from raw individual connectivity matrices, and finally letting the adversarial component drive the representation toward site invariance while preserving diagnostic utility, as evidenced by higher classification accuracy and lower site regression performance on theAB

What carries the argument

MSalNET, which uses a node information assembly mechanism to aggregate edge information from both directions into node features, a site-level feature extraction module that learns directly from individual functional connectivity data, and an adversarial learning network with a custom loss that trades off individual classification against site regression.

If this is right

Classification accuracy reaches 75.56 percent on ABIDE and 68.92 percent on ADHD-200, exceeding related algorithms.
Site regression performance drops, showing that site variability has been reduced in a data-driven way.
The most discriminative brain regions identified by the node assembly step align with those found by conventional statistical tests.
The site feature module works without any external prior information about scanning protocols or populations.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same adversarial balancing principle could be tested on other multi-site imaging modalities such as EEG or PET to check whether site effects are similarly suppressed.
If the node assembly step consistently highlights regions already known from statistics, it may offer a way to interpret deep models of brain connectivity without post-hoc explanations.
Extending the loss to include additional known confounders such as age or sex could further tighten the invariance property across datasets.

Load-bearing premise

The adversarial balance can strip site-specific information from the features without losing disorder-related signals or causing overfitting on the two datasets tested.

What would settle it

Running the same architecture on a fresh multi-site fMRI collection and finding that site regression accuracy stays high while disorder classification accuracy falls below standard baselines would falsify the claim.

read the original abstract

In open data sets of functional magnetic resonance imaging (fMRI), the heterogeneity of the data is typically attributed to a combination of factors, including differences in scanning procedures, the presence of confounding effects, and population diversities between multiple sites. These factors contribute to the diminished effectiveness of representation learning, which in turn affects the overall efficacy of subsequent classification procedures. To address these limitations, we propose a novel multi-site adversarial learning network (MSalNET) for fMRI-based mental disorder detection. Firstly, a representation learning module is introduced with a node information assembly (NIA) mechanism to better extract features from functional connectivity (FC). This mechanism aggregates edge information from both horizontal and vertical directions, effectively assembling node information. Secondly, to generalize the feature across sites, we proposed a site-level feature extraction module that can learn from individual FC data, which circumvents additional prior information. Lastly, an adversarial learning network is proposed as a means of balancing the trade-off between individual classification and site regression tasks, with the introduction of a novel loss function. The proposed method was evaluated on two multi-site fMRI datasets, i.e., Autism Brain Imaging Data Exchange (ABIDE) and ADHD-200. The results indicate that the proposed method achieves a better performance than other related algorithms with the accuracy of 75.56 and 68.92 in ABIDE and ADHD-200 datasets, respectively. Furthermore, the result of the site regression indicates that the proposed method reduces site variability from a data-driven perspective. The most discriminative brain regions revealed by NIA are consistent with statistical findings, uncovering the "black box" of deep learning to a certain extent.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

MSalNET combines node information assembly with prior-free site extraction inside an adversarial setup to cut multi-site effects in fMRI connectivity data, delivering modest accuracy gains on ABIDE and ADHD-200.

read the letter

The paper's main point is that their MSalNET architecture improves disorder classification on two multi-site fMRI datasets by pulling node features via a horizontal-vertical assembly step, learning site representations directly from individual connectivity matrices, and using adversarial training plus a custom loss to trade off classification against site regression. The reported accuracies are 75.56 on ABIDE and 68.92 on ADHD-200, with the site regression test showing lower site predictability after training. The NIA-derived regions also line up with prior statistical maps. That combination of pieces is what they flag as new, and the stress-test confirms the internal logic holds without hidden contradictions or unstated splits that would break the claims. The tables compare against baselines, which is the right thing to do here. The work is straightforward domain-adaptation applied to functional connectivity for autism and ADHD. It stays within those bounds and does not claim broader clinical impact. The gains look incremental rather than large, and the novel loss still requires weighting choices that could be sensitive to the specific data splits. Those are normal issues for this kind of paper rather than load-bearing problems. The full manuscript supplies the methods and results that the abstract alone left unclear, so the central empirical story is now verifiable. This is the kind of targeted methods paper that people working on multi-site neuroimaging pipelines would want to see. It is not reshaping the field, but the execution is clean enough that it belongs in the literature. I would send it to peer review.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes MSalNET, a multi-site adversarial learning network for fMRI-based mental disorder identification. It introduces a Node Information Assembly (NIA) mechanism to aggregate horizontal and vertical edge information from functional connectivity matrices, a site-level feature extraction module that operates on individual FC data without external priors, and an adversarial training framework with a novel loss function to balance disorder classification against site regression. On the ABIDE and ADHD-200 datasets the method is reported to reach accuracies of 75.56 and 68.92, respectively, outperforming related algorithms while also reducing site variability as measured by site-regression performance; NIA-derived regions are additionally shown to align with prior statistical findings.

Significance. If the empirical results are shown to be robust, the work addresses a practically important problem in neuroimaging: mitigating scanner- and site-induced heterogeneity without discarding disorder-relevant signal. The data-driven adversarial approach and the interpretability provided by NIA constitute modest but concrete contributions. The manuscript does not supply machine-checked proofs, open reproducible code, or parameter-free derivations, so its primary value remains empirical.

major comments (2)

[Abstract and Results] Abstract and §4 (Results): the central performance claims (accuracy 75.56 on ABIDE, 68.92 on ADHD-200, reduced site variability) are stated without any accompanying information on subject counts, cross-validation scheme, baseline methods, statistical significance tests, or error bars. These omissions directly undermine the load-bearing claim of superiority.
[§3.3] §3.3 (Adversarial Learning Network): the novel loss function is introduced to balance the classification and site-regression tasks, yet the weighting hyperparameters are listed among the free parameters and no derivation or cross-validation protocol is supplied to show they were fixed independently of the reported test metrics. This creates a moderate circularity risk for the performance numbers.

minor comments (2)

[§3.1–3.2] Notation for the NIA aggregation (horizontal vs. vertical) and the precise form of the site-level module could be clarified with an explicit equation or pseudocode block.
[Figures 3–5] Figure captions and axis labels should explicitly state whether reported accuracies are means across folds and whether error bars represent standard deviation or standard error.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive review. The comments highlight important aspects of experimental reporting and methodological transparency that we will address in the revision. Our point-by-point responses follow.

read point-by-point responses

Referee: [Abstract and Results] Abstract and §4 (Results): the central performance claims (accuracy 75.56 on ABIDE, 68.92 on ADHD-200, reduced site variability) are stated without any accompanying information on subject counts, cross-validation scheme, baseline methods, statistical significance tests, or error bars. These omissions directly undermine the load-bearing claim of superiority.

Authors: We agree that the abstract and results section would be strengthened by including these details. In the revised manuscript we will report the total number of subjects and per-site counts for both ABIDE and ADHD-200, specify the cross-validation scheme (stratified 10-fold), enumerate all baseline methods together with their accuracies, include statistical significance tests (e.g., paired t-tests or Wilcoxon tests against the strongest baselines), and add error bars or standard deviations to the reported accuracies. These additions draw on the experiments already performed and will not change the numerical results. revision: yes
Referee: [§3.3] §3.3 (Adversarial Learning Network): the novel loss function is introduced to balance the classification and site-regression tasks, yet the weighting hyperparameters are listed among the free parameters and no derivation or cross-validation protocol is supplied to show they were fixed independently of the reported test metrics. This creates a moderate circularity risk for the performance numbers.

Authors: We thank the referee for identifying this transparency gap. The weighting hyperparameters were selected via grid search on a held-out validation partition using inner cross-validation, ensuring no leakage from the final test set. In the revision we will expand §3.3 with an explicit description of the tuning protocol, the explored range, the selected values, and confirmation that selection preceded evaluation on the held-out test data. This addition will remove any appearance of circularity while preserving the reported performance figures. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper describes an empirical ML framework (MSalNET with NIA, site-level extraction, and adversarial training) whose central claims consist of reported accuracies (75.56 on ABIDE, 68.92 on ADHD-200) and site-regression outcomes obtained by running the model on the two datasets. No derivation chain reduces any claimed result to its own inputs by construction: the novel loss is presented as a balancing mechanism whose weighting parameters are chosen to produce the observed performance, but the performance numbers themselves are not redefined as the loss or vice versa. No self-citation load-bearing step, uniqueness theorem, or ansatz smuggling appears in the abstract or described architecture. The argument is therefore self-contained against external benchmarks and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 2 invented entities

The central claim rests on empirical performance of newly proposed modules whose effectiveness is not derived from first principles but asserted through reported accuracies on two datasets.

free parameters (1)

adversarial loss weighting hyperparameters
The novel loss function requires weights to trade off classification and site regression; these are typically chosen or tuned on the data.

axioms (1)

domain assumption Adversarial training can separate site-specific confounding information from disorder-relevant features in functional connectivity data.
This premise underpins the entire adversarial component and is invoked when the paper states the network balances the two tasks.

invented entities (2)

Node Information Assembly (NIA) mechanism no independent evidence
purpose: Aggregates edge information from horizontal and vertical directions to assemble node features from functional connectivity matrices.
Newly introduced component with no independent evidence outside the paper.
Site-level feature extraction module no independent evidence
purpose: Learns site-related features directly from individual FC data without additional prior information.
Newly introduced component with no independent evidence outside the paper.

pith-pipeline@v0.9.0 · 5855 in / 1560 out tokens · 79833 ms · 2026-05-22T20:56:43.130896+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

an adversarial learning network is proposed as a means of balancing the trade-off between individual classification and site regression tasks, with the introduction of a novel loss function... Lt(Xi, Yi, Ci; θE, θc) = Lc(Xi, Yi; θE, θc) + α LR(Xi, Ci)
IndisputableMonolith/Foundation/DimensionForcing.lean alexander_duality_circle_linking unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

NIA consists of two convolutional layers and one fully connected layer... horizontal convolutional kernel of the form 64@1*200... vertical kernels... 128@200*1

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.