arxiv: 2605.07026 · v1 · submitted 2026-05-07 · 🧬 q-bio.NC · cs.AI· cs.LG

Recognition: 2 theorem links

· Lean Theorem

Learning Cross-Atlas Consistent Brain Disorder Representations via Disentangled Multi-Atlas Functional Connectivity Learning

Chao Cao, Dajiang Zhu, Jing Zhang, Minheng Chen, Tianming Liu

Pith reviewed 2026-05-11 01:11 UTC · model grok-4.3

classification 🧬 q-bio.NC cs.AIcs.LG

keywords functional connectivitymulti-atlas learningdisentangled representationsbrain disordersresting-state fMRIAlzheimer's diseaseADHD

0 comments

The pith

MADCLE disentangles functional connectivity to learn cross-atlas consistent representations of brain disorders.

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

The paper introduces MADCLE, a multi-branch framework that processes functional connectivity matrices from multiple brain atlases at once. It learns atlas-specific disease-related representations and aligns their distributions to enforce consistency across parcellations. Covariate effects and atlas-dependent residuals are isolated through similarity supervision, reconstruction losses, and decorrelation constraints. This setup aims to produce more reliable embeddings for identifying disorders such as Alzheimer's and ADHD, where atlas choice otherwise introduces heterogeneity.

Core claim

MADCLE jointly encodes FC matrices derived from different brain atlases in parallel branches. Atlas-wise disease-related representations are learned and encouraged to be cross-atlas consistent through distributional alignment. Covariate-related factors receive similarity supervision, atlas-dependent residuals are modeled via atlas-specific reconstruction and decorrelation constraints, and these steps reduce leakage of non-disease or parcellation-specific information into the disease embeddings.

What carries the argument

Multi-branch disentangled representation learning that applies distributional alignment to disease factors while separating covariate and atlas-specific residuals through supervision, reconstruction, and decorrelation.

If this is right

MADCLE produces competitive or improved performance on disorder identification tasks compared with single-atlas baselines and other multi-atlas GNN or Transformer models.
Disease-related embeddings become less contaminated by parcellation-specific features.
The framework supports more stable FC-based disorder classification when multiple heterogeneous atlases are available.
Structured disentanglement offers an alternative to shallow feature fusion across atlases.

Where Pith is reading between the lines

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

The method could be applied to other neuroimaging modalities or additional clinical cohorts to test broader robustness.
In applied settings it might reduce the practical impact of choosing one particular brain atlas over another.
Direct measurement of alignment quality between disease representations from different atlases on new datasets would provide further validation.

Load-bearing premise

The combination of distributional alignment, covariate similarity supervision, atlas-specific reconstruction, and decorrelation successfully isolates disease signals without losing useful information or introducing alignment artifacts.

What would settle it

Training the model without the distributional alignment or decorrelation terms and observing whether classification accuracy on ADNI or ADHD-200 drops or cross-atlas consistency metrics fail to improve.

Figures

Figures reproduced from arXiv: 2605.07026 by Chao Cao, Dajiang Zhu, Jing Zhang, Minheng Chen, Tianming Liu.

**Figure 2.** Figure 2: Discriminative functional connections between brain regions under (a) [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

read the original abstract

Functional connectivity (FC) derived from resting-state fMRI is widely used to characterize large-scale brain network alterations in neurological and psychiatric disorders. However, FC construction critically depends on the choice of brain atlas, and different parcellations may emphasize distinct organizational features, leading to heterogeneous and sometimes inconsistent representations. Existing multi-atlas approaches partially alleviate this issue but often fuse atlas-derived features or predictions at a relatively shallow level, while single-atlas disentanglement methods do not explicitly address cross-atlas heterogeneity. We propose Multi-Atlas Disentangled Connectivity LEarning (MADCLE), a multi-branch representation learning framework that jointly encodes FC matrices derived from different brain atlases. Rather than introducing a single explicitly shared latent variable across parcellations, MADCLE learns atlas-wise disease-related representations and encourages them to be cross-atlas consistent through distributional alignment. Meanwhile, covariate-related and atlas-dependent residual factors are modeled separately using covariate similarity supervision, atlas-specific reconstruction, and decorrelation constraints, thereby reducing the leakage of non-disease and parcellation-dependent information into the disease-related embeddings. Experiments on the ADNI and ADHD-200 datasets suggest that MADCLE achieves competitive or improved performance compared with single-atlas baselines, multi-atlas GNN/Transformer models, and recent multi-atlas consistency frameworks. These results support the potential value of structured disentanglement for FC-based disorder identification under heterogeneous parcellation schemes.

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

MADCLE adds a multi-branch disentangled setup with distributional alignment on disease factors to handle atlas variability in FC disorder studies, but the claims rest on thin experimental evidence.

read the letter

The main thing to know is that this paper builds a multi-branch network for functional connectivity matrices from different atlases. Each branch produces its own disease-related embedding, which gets aligned across atlases via a distributional loss, while covariate effects and atlas residuals are pushed into separate paths with similarity supervision, per-atlas reconstruction, and decorrelation terms. That architecture is the concrete advance over prior shallow multi-atlas fusion or single-atlas disentanglement work.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces MADCLE, a multi-branch representation learning framework for functional connectivity (FC) matrices from multiple brain atlases in neurological and psychiatric disorder identification. It learns atlas-wise disease-related embeddings encouraged to be cross-atlas consistent via distributional alignment, while separately modeling covariate-related factors (via similarity supervision) and atlas-dependent residuals (via atlas-specific reconstruction and decorrelation constraints) to reduce non-disease and parcellation leakage. Experiments on the ADNI and ADHD-200 datasets are reported to show competitive or improved performance relative to single-atlas baselines, multi-atlas GNN/Transformer models, and recent multi-atlas consistency methods.

Significance. If the disentanglement successfully isolates consistent disease signals without discarding useful variance or introducing alignment artifacts, the framework could meaningfully improve robustness of FC-based disorder classification under heterogeneous parcellation schemes, a persistent issue in rs-fMRI analysis. The structured separation of factors is a principled design choice that goes beyond shallow fusion approaches.

major comments (2)

[Abstract] Abstract and Experiments: The headline claim of competitive or improved performance is presented without any reported statistical testing, error bars, cross-validation details, or ablation studies isolating the contributions of distributional alignment, covariate supervision, reconstruction, and decorrelation. This leaves the attribution of gains to the proposed disentanglement unsupported by verifiable evidence.
[Method] Method and Experiments: Distributional alignment of atlas-wise disease representations is central to the consistency claim, yet no direct checks (e.g., cross-atlas agreement on disease predictions, embedding similarity metrics for the disease factors, or analysis of retained disease variance) are described. Downstream accuracy alone cannot confirm that alignment captures shared disease mechanisms rather than marginal distributions, particularly when atlases emphasize different scales.

minor comments (1)

The abstract and method description would benefit from explicit notation for the alignment loss (e.g., MMD or adversarial) and the weighting hyperparameters of the combined objective to aid reproducibility.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment point by point below and have revised the manuscript to incorporate additional analyses and details where the concerns are valid.

read point-by-point responses

Referee: [Abstract] Abstract and Experiments: The headline claim of competitive or improved performance is presented without any reported statistical testing, error bars, cross-validation details, or ablation studies isolating the contributions of distributional alignment, covariate supervision, reconstruction, and decorrelation. This leaves the attribution of gains to the proposed disentanglement unsupported by verifiable evidence.

Authors: We agree that the original manuscript lacked explicit statistical testing, error bars, cross-validation details, and component-wise ablations, which weakens the support for attributing performance gains specifically to the disentanglement components. In the revised version, we have added 5-fold cross-validation results with standard error bars across all reported metrics, paired t-tests comparing MADCLE against baselines (with p-values), and ablation studies that isolate the effect of each term (distributional alignment, covariate similarity supervision, atlas-specific reconstruction, and decorrelation constraints). These are now included in the Experiments section with updated tables and figures, and the abstract has been revised to reference the cross-validation and statistical evaluation. This provides the requested verifiable evidence. revision: yes
Referee: [Method] Method and Experiments: Distributional alignment of atlas-wise disease representations is central to the consistency claim, yet no direct checks (e.g., cross-atlas agreement on disease predictions, embedding similarity metrics for the disease factors, or analysis of retained disease variance) are described. Downstream accuracy alone cannot confirm that alignment captures shared disease mechanisms rather than marginal distributions, particularly when atlases emphasize different scales.

Authors: We acknowledge that relying solely on downstream accuracy is insufficient to confirm the alignment isolates shared disease mechanisms. In the revised manuscript, we have added direct verification analyses in the Experiments section: (1) cross-atlas agreement on disease predictions derived from the disease-related embeddings, (2) cosine similarity and distributional distance metrics between atlas-wise disease factors, and (3) retained disease variance assessment via performance comparison with and without the alignment term. These results indicate improved cross-atlas consistency while preserving task-relevant variance, supporting that the alignment targets shared mechanisms beyond marginal distributions. We have also clarified the design rationale in the Method section to address scale differences across atlases. revision: yes

Circularity Check

0 steps flagged

No significant circularity; new multi-branch architecture and losses validated externally

full rationale

The paper introduces MADCLE as a novel multi-branch representation learning framework with distributional alignment, covariate supervision, reconstruction, and decorrelation losses to achieve cross-atlas consistent disease representations from FC matrices. The central claims rest on empirical performance comparisons against baselines on the independent ADNI and ADHD-200 datasets rather than any self-referential derivation, fitted parameter renamed as prediction, or load-bearing self-citation chain. No equations or steps in the provided description reduce the reported results to inputs by construction, and the method's outputs are tested on external data without internal tautology.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claim rests on domain assumptions about shared disease signals across atlases and the effectiveness of the proposed disentanglement losses; no new physical entities are postulated. Free parameters consist of standard deep-learning hyperparameters and loss-balancing weights that are tuned on validation data.

free parameters (2)

loss weights for alignment, reconstruction, and decorrelation terms
Chosen to balance the multiple objectives during training; typical in multi-task representation learning.
network architecture hyperparameters (depth, width, learning rate)
Standard deep learning choices optimized on held-out data.

axioms (2)

domain assumption Disease-related signals in functional connectivity are sufficiently shared across different brain atlases to permit distributional alignment.
Invoked to justify the cross-atlas consistency objective for the disease branch.
domain assumption Covariate and atlas-specific factors can be isolated via similarity supervision and reconstruction without interfering with disease representations.
Underlies the separate modeling branches and decorrelation constraints.

pith-pipeline@v0.9.0 · 5571 in / 1451 out tokens · 44171 ms · 2026-05-11T01:11:27.028361+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
MADCLE learns atlas-wise disease-related representations and encourages them to be cross-atlas consistent through distributional alignment... covariate similarity supervision, atlas-specific reconstruction, and decorrelation constraints
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear
Experiments on the ADNI and ADHD-200 datasets suggest that MADCLE achieves competitive or improved performance

Reference graph

Works this paper leans on

16 extracted references

[1]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Chen, M., Yu, X., Zhang, J., Chen, T., Cao, C., Zhuang, Y., Lyu, Y., Zhang, L., Liu, T., Zhu, D.: Core-periphery principle guided state space model for func- tional connectome classification. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 236–246. Springer (2025)

2025
[2]

Chu, Y., Wang, G., Cao, L., Qiao, L., Liu, M.: Multi-scale graph representation learningforautismidentificationwithfunctionalmri.FrontiersinNeuroinformatics 15, 802305 (2022)

2022
[3]

Human brain mapping33(8), 1914–1928 (2012)

Craddock, R.C., James, G.A., Holtzheimer III, P.E., Hu, X.P., Mayberg, H.S.: A whole brain fmri atlas generated via spatially constrained spectral clustering. Human brain mapping33(8), 1914–1928 (2012)

1914
[4]

Neuroimage31(3), 968–980 (2006)

Desikan, R.S., Ségonne, F., Fischl, B., Quinn, B.T., Dickerson, B.C., Blacker, D., Buckner, R.L., Dale, A.M., Maguire, R.P., Hyman, B.T., et al.: An automated labeling system for subdividing the human cerebral cortex on mri scans into gyral based regions of interest. Neuroimage31(3), 968–980 (2006)

2006
[5]

Cerebral cortex26(8), 3508–3526 (2016)

Fan, L., Li, H., Zhuo, J., Zhang, Y., Wang, J., Chen, L., Yang, Z., Chu, C., Xie, S., Laird, A.R., et al.: The human brainnetome atlas: a new brain atlas based on connectional architecture. Cerebral cortex26(8), 3508–3526 (2016)

2016
[6]

Proceedings of the National Academy of Sciences102(27), 9673– 9678 (2005)

Fox, M.D., Snyder, A.Z., Vincent, J.L., Corbetta, M., Van Essen, D.C., Raichle, M.E.: The human brain is intrinsically organized into dynamic, anticorrelated func- tional networks. Proceedings of the National Academy of Sciences102(27), 9673– 9678 (2005)

2005
[7]

In: 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI)

Han, K., Hu, D., Cheng, J., Liu, T., Bozoki, A., Zhu, D., Li, G.: Dual multi- atlas representation alignment for brain disorder diagnosis using morphological connectome. In: 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI). pp. 1–4. IEEE (2025)

2025
[8]

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

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

2022
[9]

NeuroImage146, 1038– 1049 (2017)

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

2017
[10]

In: International Workshop on Machine Learning in Clinical Neuroimaging

Mahler, L., Wang, Q., Steiglechner, J., Birk, F., Heczko, S., Scheffler, K., Lohmann, G.: Pretraining is all you need: A multi-atlas enhanced transformer framework for autism spectrum disorder classification. In: International Workshop on Machine Learning in Clinical Neuroimaging. pp. 123–132. Springer (2023)

2023
[11]

Human brain mapping 35(10), 5052–5070 (2014)

Min, R., Wu, G., Cheng, J., Wang, Q., Shen, D., Initiative, A.D.N.: Multi-atlas based representations for alzheimer’s disease diagnosis. Human brain mapping 35(10), 5052–5070 (2014)

2014
[12]

Neuroimage15(1), 273–289 (2002)

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

2002
[13]

Human Brain Mapping46(5), e70206 (2025) Multi-Atlas Disentangled Learning for Brain Disease Identification 11

Wu, X., Liang, C., Bustillo, J., Kochunov, P., Wen, X., Sui, J., Jiang, R., Yang, X., Fu, Z., Zhang, D., et al.: The impact of atlas parcellation on functional connectivity analysis across six psychiatric disorders. Human Brain Mapping46(5), e70206 (2025) Multi-Atlas Disentangled Learning for Brain Disease Identification 11

2025
[14]

IEEE Journal of Biomedical and Health Informatics pp

Xu, J., Lan, M., Dong, X., He, K., Zhang, W., Bian, Q., Ke, Y.: Multi-atlas brain network classification through consistency distillation and complementary informa- tion fusion. IEEE Journal of Biomedical and Health Informatics pp. 1–12 (2025)

2025
[15]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Zhang, S., Jiang, Z., Shen, X., Yu, Z., Chen, X., Zhang, X.Y., Zhou, Y.: Graph disentanglement learning for fmri analysis: Decoupling disease, covariates, and in- dividual variability. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 352–361. Springer (2025)

2025
[16]

Human brain mapping35(7), 2911–2923 (2014)

Zhu, D., Li, K., Terry, D.P., Puente, A.N., Wang, L., Shen, D., Miller, L.S., Liu, T.: Connectome-scale assessments of structural and functional connectivity in mci. Human brain mapping35(7), 2911–2923 (2014)

2014