pith. sign in

arxiv: 1907.01288 · v1 · pith:CWVLHPYTnew · submitted 2019-07-02 · 🧬 q-bio.NC · cs.LG· eess.IV· stat.ML

Simple 1-D Convolutional Networks for Resting-State fMRI Based Classification in Autism

Ahmed El Gazzar , Leonardo Cerliani , Guido van Wingen , Rajat Mani Thomas This is my paper

Pith reviewed 2026-05-25 10:49 UTC · model grok-4.3

classification 🧬 q-bio.NC cs.LGeess.IVstat.ML
keywords resting-state fMRIautism spectrum disorder1-D convolutional networksclassificationneuroimagingminimal preprocessing
0
0 comments X

The pith

A simple spatial subsampling of rsfMRI data lets a basic 1-D convolutional network match state-of-the-art autism classification performance.

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

The paper establishes that transforming rsfMRI volumes by subsampling their spatial extent while keeping all temporal dynamics intact allows a very simple 1-D convolutional network to reach classification accuracy for autism spectrum disorder that is on par with far more elaborate models. This holds even with minimal preprocessing and rapid training times. A sympathetic reader would care because the result indicates that diagnostically useful information in resting-state connectivity can survive aggressive spatial reduction, lowering the computational barrier to applying deep learning in psychiatric neuroimaging. The central object carrying the argument is the subsampled time-series representation itself.

Core claim

A very simple transformation of the rsfMRI images that captures all of the temporal dynamics of the signal but sub-samples its spatial extent allows a very simple 1-D convolutional network to perform at par with the state-of-the-art on the classification of Autism spectrum disorders while requiring minimal preprocessing and fast training.

What carries the argument

the spatially subsampled rsfMRI time-series representation passed to a 1-D convolutional network

If this is right

  • Training requires only standard hardware and finishes rapidly.
  • Extensive preprocessing pipelines become unnecessary.
  • Model complexity can be kept low without sacrificing accuracy on this task.
  • Temporal dynamics alone suffice once spatial resolution is reduced in this manner.

Where Pith is reading between the lines

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

  • The same subsampling step could be tested on other psychiatric or neurological classification tasks that use rsfMRI.
  • If the approach generalizes, it would imply that full spatial resolution is often redundant for connectivity-based diagnosis.
  • One could measure exactly how coarse the spatial grid can become before accuracy drops, providing a practical limit for future pipelines.

Load-bearing premise

The spatial subsampling step preserves all diagnostically relevant information present in the original high-dimensional rsfMRI volume.

What would settle it

A controlled experiment showing that the same 1-D network trained on the full-resolution rsfMRI volumes achieves materially higher accuracy than on the subsampled version would falsify the claim that the subsampling step retains all necessary information.

Figures

Figures reproduced from arXiv: 1907.01288 by Ahmed El Gazzar, Guido van Wingen, Leonardo Cerliani, Rajat Mani Thomas.

Figure 2
Figure 2. Figure 2: Leave site out validation results using AAL, Harvard Oxford, Schaefer [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Increasing mean TSNR across sites vs leave site out validation [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
read the original abstract

Deep learning methods are increasingly being used with neuroimaging data like structural and function magnetic resonance imaging (MRI) to predict the diagnosis of neuropsychiatric and neurological disorders. For psychiatric disorders in particular, it is believed that one of the most promising modality is the resting-state functional MRI (rsfMRI), which captures the intrinsic connectivity between regions in the brain. Because rsfMRI data points are inherently high-dimensional (~1M), it is impossible to process the entire input in its raw form. In this paper, we propose a very simple transformation of the rsfMRI images that captures all of the temporal dynamics of the signal but sub-samples its spatial extent. As a result, we use a very simple 1-D convolutional network which is fast to train, requires minimal preprocessing and performs at par with the state-of-the-art on the classification of Autism spectrum disorders.

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.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a spatial subsampling transformation of rsfMRI volumes that retains full temporal dynamics while reducing spatial dimensionality, allowing a simple 1-D convolutional network to classify autism spectrum disorder (ASD) versus controls. The central claim is that this approach requires minimal preprocessing and achieves performance at parity with state-of-the-art methods on ASD classification.

Significance. If the performance parity is robustly validated, the work would offer a practical simplification for rsfMRI-based psychiatric classification by lowering computational demands and preprocessing complexity. The emphasis on a minimal architecture is a strength that could aid reproducibility and accessibility if supported by code release and detailed validation protocols.

major comments (2)
  1. [Abstract] Abstract, paragraph 3: the assertion that spatial subsampling 'captures all of the temporal dynamics' while preserving diagnostically relevant information is load-bearing for the parity claim, yet the manuscript provides no ablation study comparing the subsampled representation against models trained on the original full-resolution (~1M-dimensional) volumes. Without this comparison, it remains unclear whether the reported performance reflects sufficiency of the reduced input or merely the capacity of the 1D CNN on the chosen dataset.
  2. [Methods/Results] Methods/Results sections (inferred from abstract description): no quantitative details are supplied on dataset sizes, number of subjects, cross-validation scheme, or statistical testing (e.g., permutation tests or confidence intervals) used to establish 'at par with SOTA'. These omissions prevent evaluation of whether the parity claim is statistically supported or sensitive to particular data splits.
minor comments (2)
  1. [Abstract] Abstract: 'function magnetic resonance imaging' should read 'functional magnetic resonance imaging'.
  2. [Abstract] Abstract: 'at par with' is nonstandard; 'on par with' is the conventional phrasing.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments. We address each major comment below with the strongest honest defense of the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract, paragraph 3: the assertion that spatial subsampling 'captures all of the temporal dynamics' while preserving diagnostically relevant information is load-bearing for the parity claim, yet the manuscript provides no ablation study comparing the subsampled representation against models trained on the original full-resolution (~1M-dimensional) volumes. Without this comparison, it remains unclear whether the reported performance reflects sufficiency of the reduced input or merely the capacity of the 1D CNN on the chosen dataset.

    Authors: The spatial subsampling is constructed to retain every time point from the selected voxels, so all temporal dynamics are captured by design; only spatial extent is reduced. Full-resolution (~1M-dimensional) inputs cannot be processed by the same simple 1D CNN without changing the architecture entirely, which would negate the method's purpose of enabling minimal models. The parity with SOTA on the subsampled data therefore demonstrates that diagnostically relevant information is retained. We will add an explicit statement of this rationale to the abstract and discussion. revision: partial

  2. Referee: [Methods/Results] Methods/Results sections (inferred from abstract description): no quantitative details are supplied on dataset sizes, number of subjects, cross-validation scheme, or statistical testing (e.g., permutation tests or confidence intervals) used to establish 'at par with SOTA'. These omissions prevent evaluation of whether the parity claim is statistically supported or sensitive to particular data splits.

    Authors: The full manuscript reports the ABIDE dataset composition, subject counts, cross-validation scheme, and statistical comparisons to SOTA. To prevent any ambiguity we will add a concise summary of these details to the abstract and ensure they are explicitly referenced in the results. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical performance claim on subsampled rsfMRI

full rationale

The paper proposes a spatial subsampling of rsfMRI volumes that retains temporal dynamics, then applies a 1D CNN for ASD classification and reports empirical parity with SOTA. This is a standard ML experiment with no derivation chain, no fitted parameters renamed as predictions, and no self-citation load-bearing the central result. The information-preservation assumption is tested (or asserted) via downstream accuracy rather than being true by construction. No equations or self-referential reductions appear in the provided text.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities; the transformation step is described only at the level of 'captures all temporal dynamics but sub-samples spatial extent' without further specification.

pith-pipeline@v0.9.0 · 5696 in / 1100 out tokens · 16880 ms · 2026-05-25T10:49:24.118181+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages · 2 internal anchors

  1. [1]

    Using deep learning to in- vestigate the neuroimaging correlates of psychiatric and neurological disorders: Methods and applications,

    Vieira S, Pinaya WH, and Mechelli A, “Using deep learning to in- vestigate the neuroimaging correlates of psychiatric and neurological disorders: Methods and applications,” Neuroscience & Biobehavioral Reviews. 2017 Mar 1;74:58-75

    work page 2017

  2. [2]

    Frequencies contributing to functional connectivity in the cerebral cortex in

    Cordes, D, Haughton, VM, Arfanakis, K, Carew, JD, Turski, PA, Moritz, CH, Quigley, MA, Meyerand, ME, “Frequencies contributing to functional connectivity in the cerebral cortex in ”resting-state” ” data. AJNR Am J Neuroradiol, 2001, 22:13261333

    work page 2001

  3. [3]

    Improved optimization for the robust and accurate linear registration and motion correction of brain images,

    Jenkinson, M, Bannister, P, Brady, M, Smith, S, “Improved optimization for the robust and accurate linear registration and motion correction of brain images,” Neuroimage, 2002, 17:825841

    work page 2002

  4. [4]

    ICA-AROMA: A robust ICA-based strategy for removing motion artifacts from fMRI data,

    Pruim, RH, Mennes, M, van Rooij, D, Llera, A, Buitelaar, JK, Beck- mann, CF, “ICA-AROMA: A robust ICA-based strategy for removing motion artifacts from fMRI data,” Neuroimage, 2015, 112:267277

    work page 2015

  5. [5]

    Fast robust automated brain extraction,

    Smith, SM, “Fast robust automated brain extraction,” Hum Brain Mapp, 2002, 17:143155

    work page 2002

  6. [6]

    The influence of head motion on intrinsic functional connectivity MRI,

    Van Dijk, KR, Sabuncu, MR, Buckner, RL, “The influence of head motion on intrinsic functional connectivity MRI,” Neuroimage, 2012, 59:431438

    work page 2012

  7. [7]

    On the origin of respiratory artifacts in BOLD-EPI of the human brain,

    Windischberger, C, Langenberger, H, Sycha, T, Tschernko, EM, Fuchsjager-Mayerl, G, Schmetterer, L, Moser, E, “On the origin of respiratory artifacts in BOLD-EPI of the human brain,” Magn Reson Imaging, 2002, 20:575582

    work page 2002

  8. [8]

    Bayesian analysis of neuroimaging data in FSL,

    Woolrich, MW, Jbabdi, S, Patenaude, B, Chappell, M, Makni, S, Behrens, T, Beckmann, C, Jenkinson, M, Smith, SM, “Bayesian analysis of neuroimaging data in FSL,” Neuroimage, 2009, 45:S17386

    work page 2009

  9. [9]

    Segmentation of brain MR images through a hidden Markov random field model and the expectation- maximization algorithm,

    Zhang, Y , Brady, M, Smith, S, “Segmentation of brain MR images through a hidden Markov random field model and the expectation- maximization algorithm,” IEEE Trans Med Imaging, 2001 20:4557

    work page 2001

  10. [10]

    Automated anatomical labeling of activations in SPM using a macroscopic anatom- ical parcellation of the MNI MRI single-subject brain,

    N. Tzourio-Mazoyer, B. Landeau, D. Papathanassiou, F. Crivello, O. Etard, N. Delcroix, Bernard Mazoyer, and M. Joliot, “Automated anatomical labeling of activations in SPM using a macroscopic anatom- ical parcellation of the MNI MRI single-subject brain,” NeuroImage, 2002, 15: 273-289

    work page 2002

  11. [11]

    Local-Global parcellation of the human cerebral cortex from intrinsic functional connectivity MRI, Cerebral Cortex, 2018, 29:3095-3114

    Schaefer A, Kong R, Gordon EM, Laumann TO, Zuo XN, Holmes AJ, Eickhoff SB, and Yeo BTT. Local-Global parcellation of the human cerebral cortex from intrinsic functional connectivity MRI, Cerebral Cortex, 2018, 29:3095-3114

    work page 2018

  12. [12]

    Decreased volume of left and total anterior insular lobule in schizophrenia,

    Makris N, Goldstein JM, Kennedy D, Hodge SM, Caviness VS, Faraone SV , Tsuang MT, Seidman LJ, “Decreased volume of left and total anterior insular lobule in schizophrenia,” Schizophr Res. 2006 Apr;83(2- 3):155-71

    work page 2006

  13. [13]

    Intrinsic Functional Connectivity as a Tool for Human Connectomics: Theory, Properties, and Optimization,

    Van Dijk, Koene R. A., Trey Hedden, Archana Venkataraman, Karleyton C. Evans, Sara W. Lazar, and Randy L. Buckner, “Intrinsic Functional Connectivity as a Tool for Human Connectomics: Theory, Properties, and Optimization,” Journal of Neurophysiology, 2010, 103 (1): 297321

    work page 2010

  14. [14]

    Deriving Reproducible Biomarkers from Multi-Site Resting-State Data: An Autism-Based Example,

    Abraham, Alexandre, Michael P. Milham, Adriana Di Martino, R. Cameron Craddock, Dimitris Samaras, Bertrand Thirion, and Gael Varoquaux. 2017. “Deriving Reproducible Biomarkers from Multi-Site Resting-State Data: An Autism-Based Example,” NeuroImage, 2017 147 (February): 73645

    work page 2017

  15. [15]

    Bootstrapping Graph Convolutional Neural Networks for Autism Spectrum Disorder Classification

    Anirudh, Rushil, and Jayaraman J. Thiagarajan. “Bootstrapping Graph Convolutional Neural Networks for Autism Spectrum Disorder Classi- fication,” 2017, arXiv [stat.ML]. arXiv. http://arxiv.org/abs/1704.07487

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  16. [16]

    The Autism Brain Imaging Data Exchange: Towards a Large-Scale Evaluation of the Intrinsic Brain Architecture in Autism,

    Di Martino, A., C. G. Yan, Q. Li, E. Denio, F. X. Castellanos, K. Alaerts, J. S. Anderson, et al, “The Autism Brain Imaging Data Exchange: Towards a Large-Scale Evaluation of the Intrinsic Brain Architecture in Autism,” 2013, Molecular Psychiatry 19: 65967

    work page 2013

  17. [17]

    Identification of Autism Spectrum Disorder Using Deep Learning and the ABIDE Dataset,

    Heinsfeld, Anibal Slon, Alexandre Rosa Franco, R. Cameron Craddock, Augusto Buchweitz, and Felipe Meneguzzi, “Identification of Autism Spectrum Disorder Using Deep Learning and the ABIDE Dataset,” 2018, NeuroImage. Clinical 17: 1623

    work page 2018

  18. [18]

    3D Convolutional Neural Networks for Classification of Functional Connectomes

    Khosla, Meenakshi, Keith Jamison, Amy Kuceyeski, and Mert Sabuncu, “3D Convolutional Neural Networks for Classification of Functional Connectomes, ” 2018, arXiv [cs.CV]. arXiv. http://arxiv.org/abs/1806.04209

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  19. [19]

    Distance Metric Learning Using Graph Convolutional Networks: Application to Functional Brain Networks,

    Ktena, Sofia Ira, Sarah Parisot, Enzo Ferrante, Martin Rajchl, Matthew Lee, Ben Glocker, and Daniel Rueckert, “Distance Metric Learning Using Graph Convolutional Networks: Application to Functional Brain Networks,” 2017, In Medical Image Computing and Computer Assisted Intervention MICCAI 2017, edited by Maxime Descoteaux, Lena Maier-Hein, Alfred Franz, Pi...

    work page 2017

  20. [20]

    Disease Prediction Us- ing Graph Convolutional Networks: Application to Autism Spectrum Disorder and Alzheimers Disease,

    Parisot, Sarah, Sofia Ira Ktena, Enzo Ferrante, Matthew Lee, Ricardo Guerrero, Ben Glocker, and Daniel Rueckert, “Disease Prediction Us- ing Graph Convolutional Networks: Application to Autism Spectrum Disorder and Alzheimers Disease,” 2018, Medical Image Analysis 48 (August): 11730

    work page 2018

  21. [21]

    Functional connectivity in resting-state fMRI: is linear correlation sufficient?

    Hlinka, J., Palus, M. and Vejmelka, M. and Mantini, D. and Corbetta, M., “Functional connectivity in resting-state fMRI: is linear correlation sufficient?” 2011, NeuroImage 54(3): 2218-2225

    work page 2011