Alzheimers Disease Classification in Functional MRI With 4D Joint Temporal-Spatial Kernels in Novel 4D CNN Model

Javier Salazar Cavazos; Scott Peltier

REVIEW 4 major objections 6 minor 7 references

Alzheimers Disease Classification in Functional MRI With 4D Joint Temporal-Spatial Kernels in Novel 4D CNN Model

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A 4D CNN with joint temporal-spatial kernels predicts Alzheimer's diagnosis from resting-state fMRI more accurately than 3D CNN and 3D CNN+LSTM baselines.

desk verdict Plausible 4D CNN architecture for AD classification, but the central claim is unsupported because Table I is empty and no statistics are reported. read the letter →

arxiv 2506.02060 v1 pith:NWU6UI5U submitted 2025-06-01 eess.IV cs.CV

Javier Salazar Cavazos , Scott Peltier This is my paper

classification eess.IVcs.CV

keywords Alzheimer'sdiseaseresting-statefMRI4DCNNtemporal-spatialkernelsdeeplearningclassificationfunctionaldiagnosis

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 treating functional MRI as a 4D volume (x, y, z, time) and convolving with joint temporal-spatial kernels extracts temporal dynamics that purely spatial 3D models miss, and that this improves Alzheimer's disease diagnosis from resting-state fMRI. It reports that a custom 4D CNN outperforms two baselines: a 3D ConvNeXt treating time samples as channels, and a 3D CNN+LSTM hybrid that separates spatial feature extraction from temporal modeling. The work is presented as the first application of a 4D CNN to Alzheimer's diagnosis with rs-fMRI. If the result holds, this points to a class of models that capture both where and when BOLD signals change, with implications for earlier detection and a template for other 4D neuroimaging tasks.

What carries the argument

The central object is the 4D convolution: a kernel with dimensions (x, y, z, time) applied directly to the fMRI volume, implemented as custom Conv4D layers. It is supported by a 1-1-3-1 downsampling architecture with 1024 output channels and global average pooling, weighted cross-entropy loss to handle class imbalance, and a circular time-shifting augmentation that enforces time-invariance. The 4D kernel is what carries the argument: unlike a 3D CNN that collapses time into channels or a hybrid that pools spatial features before feeding an LSTM, it learns joint spatial-temporal filters in one pass.

What would settle it

Run a subject-level nested cross-validation on a larger ADNI cohort, averaging over multiple random seeds, and compute confidence intervals for accuracy; if the 4D CNN's accuracy falls within the baselines' intervals or is not consistently higher across folds, the paper's central claim is refuted.

Watch

Extended reading notes

Core claim

On the ADNI dataset, the paper's custom Conv4D network — four downsampling stages in a 1-1-3-1 configuration ending in 1024 globally pooled features — classifies cognitively normal, mild cognitive impairment, and Alzheimer's disease from rs-fMRI more accurately than a 3D CNN that treats time as input channels and a 3D CNN+LSTM hybrid that models temporal dynamics separately. The 4D kernels in the first layer show interpretable low-level features such as derivatives and weighted averages over time, and Grad-CAM++ saliency maps highlight the hippocampus, cerebellum, and prefrontal cortex. The paper claims this demonstrates that joint temporal-spatial kernels, rather than separate spatial and temporal modules, are the effective way to use the full 4D fMRI signal for diagnosis.

Load-bearing premise

The central claim assumes the balanced test set of 50 samples per class reflects subject-level generalization: the number of unique subjects is not reported, and if the advantage over the 3D baselines is random variation or subject-specific artifact in this small set, the conclusion collapses.

Editorial extensions

If this is right

The 4D CNN's diagnostic accuracy advantage over both baselines suggests that joint temporal-spatial kernels are better suited to rs-fMRI than treating time as a channel dimension or modeling it separately with an LSTM.
The first-layer kernels exhibit interpretable derivative and averaging structure, indicating that the model learns meaningful temporal dynamics rather than purely spatial textures.
Grad-CAM++ saliency maps identify hippocampus, cerebellum, and prefrontal cortex as the regions driving diagnosis, consistent with known AD-relevant networks.
Because the model consumes raw preprocessed 4D volumes, it could be applied to task-based fMRI and to regression tasks such as cognitive scoring without architectural change.

Reading between the lines

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

If the improvement over the 3D+LSTM hybrid is stable across subjects and seeds, it would imply that simultaneous spatio-temporal filtering is more sample-efficient than separate spatial-then-temporal processing, potentially transferring to other clinical time-series imaging tasks.
The circular time-shifting augmentation is a cheap way to impose temporal invariance that could be adopted by other 4D neuroimaging models, including those outside Alzheimer's research.
The paper's 'pseudo-subject' augmentation treats each session as independent; a stricter test would verify that the 4D model's edge survives when multiple sessions from the same individual are held out at the subject level.
A natural extension is to test whether the same 4D kernels are useful for predicting disease progression or for classifying other dementias, which would indicate the temporal features are general rather than specific to AD.

Share X Bluesky LinkedIn Reddit HN

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

Short version: this is a one-page ISMRM abstract proposing a 4D CNN for Alzheimer's classification from resting-state fMRI, and the central claim that it outperforms 3D and 3D+LSTM baselines is currently unverifiable because Table I is empty. The method itself is coherent and the application is genuinely new; the evidence is not.

What is actually new: first use of 4D spatio-temporal convolutions for AD diagnosis in rs-fMRI. The authors are honest that 4D CNNs exist for cardiac segmentation (ref 7) and autism (ref 8); they extend the approach to a new disease and add a careful comparison against a 3D CNN and a 3D+LSTM hybrid. The preprocessing pipeline is standard and solid: BIDS, fMRIPrep, bandpass filtering, z-scoring, and a subject-disjoint train/test split. They also try to deal with class imbalance via inverse-frequency weighting and circular time shifting, which shows methodological awareness.

The soft spots are central. The Results section says "better predicted patient diagnosis" but Table I contains only a caption, no numbers. No accuracy, sensitivity, or specificity values appear anywhere in the abstract. Without them, the comparison is an assertion. There are also no confidence intervals or significance tests, so even if numbers were present, we couldn't tell whether the improvement is real or noise. The pseudo-subject augmentation is a secondary concern: the split is subject-disjoint, but if multiple sessions per subject land in the test set, the effective number of independent test subjects could be well below 50 per class, which would weaken any statistical claim. The abstract does not report unique subject counts.

Who it's for: someone working on deep learning for neuroimaging might find the architecture idea worth a quick look, but the manuscript as submitted does not provide enough evidence to act on. If the authors submit a full paper with actual results and subject-level statistics, it would deserve a thorough review. As an abstract, the central claim cannot be evaluated.

Recommendation: I would not send this to peer review in its current form; the missing table is a desk-reject-level gap. If the full paper contains the numbers and analyses, that would be a different conversation.

Referee Report

4 major / 6 minor

Summary. The manuscript proposes a 4D convolutional neural network with joint temporal-spatial kernels for classifying Alzheimer's disease from resting-state fMRI, and compares it against a 3D CNN (time-as-channels) and a 3D CNN+LSTM hybrid on ADNI data. The authors report that the 4D CNN 'better predicted patient diagnosis compared to other models' and that this is the first application of a 4D CNN to AD diagnosis from rs-fMRI. The paper includes preprocessing details, class-balanced test-set design, and interpretability analyses (kernel inspection and Grad-CAM++). However, the central quantitative claim is not supported in the submitted text because Table I, which is the sole source of the claimed comparison, contains no numerical results.

Significance. If the claimed result were substantiated, the contribution would be a modest but potentially useful application: 4D spatiotemporal convolutions on raw rs-fMRI for AD classification, with interpretability evidence pointing to known AD-relevant regions. The manuscript has several strengths: the train/test split is explicitly subject-disjoint, the test set is balanced, preprocessing follows standard fMRIPrep-based practices, and the interpretability analyses (kernel visualization and Grad-CAM++) are appropriate for a neuroimaging audience. The main weakness is that the paper's central claim—that the 4D CNN outperforms the baselines—is currently unverifiable because the results table is empty and no statistical testing is reported. The evaluation also rests on a 'pseudo-subject' augmentation whose subject-level sample size is not disclosed, which could materially affect the reliability of the stated performance.

major comments (4)

[Results, Table I] The central claim that 'The 4D CNN model better predicted patient diagnosis compared to other models' is supported only by Table I, yet Table I contains no numerical entries. The abstract and Discussion restate the claim, but no accuracy, sensitivity, or specificity values are given for any model or class setting. This is a load-bearing omission: without the quantitative comparison, the manuscript provides no evidence for the claimed superiority. Please provide the full table with per-class and multi-class metrics for all three models.
[Methods, 'pseudo-subject' augmentation] The Methods state that 'considering each session as an independent pseudo-subject' was used to augment the dataset, and that scans from the same individual were assigned exclusively to one set. This subject-disjoint design is commendable, but the number of unique subjects per class is not reported. The reported counts (CN 602/50, MCI 210/50, DAT 147/50) are session/sample counts; if multiple sessions per individual are present, the effective test size in terms of unique subjects could be considerably smaller than 50 per class, weakening the reliability of the reported accuracies and any significance test. Please report the number of unique subjects and the distribution of sessions per subject.
[Results, comparison and significance] No confidence intervals, standard deviations, or statistical significance tests accompany the claimed performance differences. With a small test set (at most 50 samples per class, and possibly fewer unique subjects), the observed difference between the 4D CNN and the baselines could arise from random variation. The authors should report paired statistical comparisons (e.g., McNemar's test or bootstrapped confidence intervals) for the accuracy and sensitivity/specificity differences across the three models.
[Discussion & Conclusion] The conclusion that the model 'outperformed other methods' is not independently supported by any presented numerical evidence; it depends entirely on the missing Table I. Additionally, the claim of 'earlier detection' in the impact statement goes beyond the cross-sectional classification setup in this study, since no longitudinal or early-stage prediction experiment is reported. Please temper the conclusion to match the actual experimental scope unless longitudinal evidence is added.

minor comments (6)

[Title and Abstract] 'Alzheimers' should be 'Alzheimer's' throughout the title, abstract, and main text.
[Methods, data distribution] There is a stray closing brace in 'CN (602/50), MCI (210/50), and DAT (147/50)}' that should be removed.
[References] Reference [6] is truncated and incomplete; the full author list and publication details should be provided.
[Methods, implementation details] The architecture descriptions reference figures but do not specify kernel sizes, number of channels per stage, stride, or the LSTM hidden dimension; these details are needed for reproducibility and should be added or made available in code.
[Figure 3 caption] The caption mentions 'first layer channels (C=128)' while the architecture description in Figure 1 refers to a final output of 1024 channels; please clarify the channel counts at each stage to avoid confusion.
[Availability] The manuscript does not state whether code or the trained models will be made publicly available; given the reproducibility expectations for deep-learning neuroimaging papers, a data/code availability statement should be added.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical model comparison with a subject-disjoint train/test split; the missing Table I is a reporting gap, not a circular step.

full rationale

This paper is an empirical supervised-learning comparison of three deep-learning architectures on ADNI rs-fMRI data. The central claim, that the 4D CNN better predicted patient diagnosis, is an experimental result on a held-out test set, not a derivation. The train/test split is explicitly subject-disjoint ('scans from the same individual were assigned exclusively to one set'), so test-set performance is not a fitted quantity and the test labels are not used in training. The loss weights, optimizer, and architecture choices are standard engineering decisions, not parameters fitted to the test set. The cited prior works on 4D CNNs (references 7 and 8) are external applications of similar ideas and are not used as load-bearing justification for the current model's superiority; the manuscript states 'not entirely unprecedented' and claims a first application to AD, which is a novelty claim, not a circular proof. No 'uniqueness theorem' is invoked, and no self-citation chain supports the result. The main weakness is that Table I, which would contain the quantitative results, is empty in the provided text, and no confidence intervals or significance tests are reported. This is a verifiability and completeness issue, not circularity. Therefore, the paper does not reduce to its own inputs by definition or by statistical construction.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on standard domain assumptions about rs-fMRI biomarkers and on unverified details of model hyperparameters and evaluation procedure.

free parameters (1)

Model hyperparameters (kernel sizes, strides, channels, learning rate, weight decay, etc.) = Not reported
The 4D CNN architecture is described only qualitatively (1-1-3-1 downsampling stages, 1024 channels); without exact values, the model cannot be replicated and the results may be sensitive to these choices.

assumptions (3)

domain assumption BOLD signal variations in rs-fMRI are informative biomarkers for Alzheimer's disease
The introduction cites refs 1-2 to justify using rs-fMRI, but this relationship is an unproved background assumption.
domain assumption The ADNI dataset labels are correct and the train/test split prevents subject-level leakage
They state scans from the same individual were assigned exclusively to one set, but no unique subject counts are provided, so the effectiveness of the split cannot be verified.
domain assumption Treating each session as an independent pseudo-subject does not introduce bias
The method section describes this augmentation but provides no analysis of how it affects the distribution or the independence of samples.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Alzheimers Disease Classification in Functional MRI With 4D Joint Temporal-Spatial Kernels in Novel 4D CNN Model." pith.science (2026). https://pith.science/paper/NWU6UI5U

@misc{pith2026250602060,
  author       = {Pith},
  title        = {Pith review of: Alzheimers Disease Classification in Functional MRI With 4D Joint Temporal-Spatial Kernels in Novel 4D CNN Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NWU6UI5U}},
  note         = {Machine review of arXiv:2506.02060}
}

read the original abstract

Previous works in the literature apply 3D spatial-only models on 4D functional MRI data leading to possible sub-par feature extraction to be used for downstream tasks like classification. In this work, we aim to develop a novel 4D convolution network to extract 4D joint temporal-spatial kernels that not only learn spatial information but in addition also capture temporal dynamics. Experimental results show promising performance in capturing spatial-temporal data in functional MRI compared to 3D models. The 4D CNN model improves Alzheimers disease diagnosis for rs-fMRI data, enabling earlier detection and better interventions. Future research could explore task-based fMRI applications and regression tasks, enhancing understanding of cognitive performance and disease progression.

Figures

Figures reproduced from arXiv: 2506.02060 by the authors.

**Figure 1.** For the 3D CNN + LSTM model, spatial features for each time sample were separately extracted and globally averaged. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗

**Figure 1.** Proposed 4D CNN model architecture, consisting of four downsampling stages in a 1-1-3-1 configuration. The final stage outputs 1024 channels that are globally average pooled to yield 1024 features for the entire 4D scan [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Architecture of the hybrid 3D CNN + LSTM model. Each 3D time sample is processed individually by the CNN, and the resulting features are aggregated into matrix S. The LSTM module captures temporal dynamics between time samples for classification purposes. Table I. Comparative results for the three approaches to handling the time dimension in raw 4D fMRI data. Accuracy, sensitivity, and specificity are reported for v… view at source ↗

Figures from the paper (2 more)

**Figure 3.** Figure 3: Temporal kernels from random spatial kernel locations for first layer channels (C=128). Only a subset of the total channels are shown for illustration simplicity. Moreover, only a few examples per filter are shown. The proposed model in the first layer extracts low-lev…

**Figure 4.** Figure 4: Model interpretability figure using the Grad-CAM++ method. Left image consists of the BOLD response at the hippocampus for a DAT diagnosed subject and corresponding Grad-CAM saliency signal over time. Right image consists of spatial Grad-CAM maps for a fixed time sampl…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

7 extracted references · 6 canonical work pages

[7]

& Roth, H

Myronenko, A., Yang, D., Buch, V., Xu, D., Ihsani, A., Doyle, S., ... & Roth, H. (2020). 4D CNN for semantic segmentation of cardiac volumetric sequences. In Statistical Atlases and Computational Models of the Heart. Multi-Sequence CMR Segmentation, CRT-EPiggy and LV Full Quantification Challenges: 10th International Workshop, STACOM 2019, Held in Conjunc...

work page 2020
[8]

Bengs, M., Gessert, N., & Schlaefer, A. (2020). 4d spatio-temporal deep learning with 4d fmri data for autism spectrum disorder classification. arXiv preprint arXiv:2004.10165. 9. Mueller, S. G., Weiner, M. W., Thal, L. J., Petersen, R. C., Jack, C., Jagust, W., ... & Beckett, L. (2005). The Alzheimer’s disease neuroimaging initiative. Neuroimaging Clinic...

work page Pith review arXiv 2020
[10]

J., Auer, T., Calhoun, V

Gorgolewski, K. J., Auer, T., Calhoun, V. D., Craddock, R. C., Das, S., Duff, E. P., ... & Poldrack, R. A. (2016). The brain imaging data structure, a format for organizing and describing outputs of neuroimaging experiments. Scientific data, 3(1), 1-9

work page 2016
[11]

J., Blair, R

Esteban, O., Markiewicz, C. J., Blair, R. W., Moodie, C. A., Isik, A. I., Erramuzpe, A., ... & Gorgolewski, K. J. (2019). fMRIPrep: a robust preprocessing pipeline for functional MRI. Nature methods, 16(1), 111-116. 12 . Ciric, R., Thompson, W. H., Lorenz, R., Goncalves, M., MacNicol, E. E., Markiewicz, C. J., ... & Esteban, O. (2022). TemplateFlow: FAIR-...

work page 2019
[14]

J., Hanke, M., Côté, M.-A., Cipollini, B., Papadopoulos Orfanos, D., McCarthy, P., Jarecka, D., Cheng, C

Brett, M., Markiewicz, C. J., Hanke, M., Côté, M.-A., Cipollini, B., Papadopoulos Orfanos, D., McCarthy, P., Jarecka, D., Cheng, C. P., Larson, E., Halchenko, Y. O., Cottaar, M., Ghosh, S., Wassermann, D., Gerhard, S., Lee, G. R., Baratz, Z., Moloney, B., Wang, H.-T., … freec84. (2024). nipy/nibabel: 5.3.1 (5.3.1). Zenodo. https://doi.org/10.5281/zenodo.13936989

work page doi:10.5281/zenodo.13936989 2024
[15]

Kingma, D. P. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

arXiv 2014
[16]

R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D

Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2017). Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision (pp. 618-626). Figures Figure 1. Proposed 4D CNN model architecture, consisting of four downsampling stages in a 1-1-...

work page 2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [7]

& Roth, H

Myronenko, A., Yang, D., Buch, V., Xu, D., Ihsani, A., Doyle, S., ... & Roth, H. (2020). 4D CNN for semantic segmentation of cardiac volumetric sequences. In Statistical Atlases and Computational Models of the Heart. Multi-Sequence CMR Segmentation, CRT-EPiggy and LV Full Quantification Challenges: 10th International Workshop, STACOM 2019, Held in Conjunc...

work page 2020

[2] [8]

Bengs, M., Gessert, N., & Schlaefer, A. (2020). 4d spatio-temporal deep learning with 4d fmri data for autism spectrum disorder classification. arXiv preprint arXiv:2004.10165. 9. Mueller, S. G., Weiner, M. W., Thal, L. J., Petersen, R. C., Jack, C., Jagust, W., ... & Beckett, L. (2005). The Alzheimer’s disease neuroimaging initiative. Neuroimaging Clinic...

work page Pith review arXiv 2020

[3] [10]

J., Auer, T., Calhoun, V

Gorgolewski, K. J., Auer, T., Calhoun, V. D., Craddock, R. C., Das, S., Duff, E. P., ... & Poldrack, R. A. (2016). The brain imaging data structure, a format for organizing and describing outputs of neuroimaging experiments. Scientific data, 3(1), 1-9

work page 2016

[4] [11]

J., Blair, R

Esteban, O., Markiewicz, C. J., Blair, R. W., Moodie, C. A., Isik, A. I., Erramuzpe, A., ... & Gorgolewski, K. J. (2019). fMRIPrep: a robust preprocessing pipeline for functional MRI. Nature methods, 16(1), 111-116. 12 . Ciric, R., Thompson, W. H., Lorenz, R., Goncalves, M., MacNicol, E. E., Markiewicz, C. J., ... & Esteban, O. (2022). TemplateFlow: FAIR-...

work page 2019

[5] [14]

J., Hanke, M., Côté, M.-A., Cipollini, B., Papadopoulos Orfanos, D., McCarthy, P., Jarecka, D., Cheng, C

Brett, M., Markiewicz, C. J., Hanke, M., Côté, M.-A., Cipollini, B., Papadopoulos Orfanos, D., McCarthy, P., Jarecka, D., Cheng, C. P., Larson, E., Halchenko, Y. O., Cottaar, M., Ghosh, S., Wassermann, D., Gerhard, S., Lee, G. R., Baratz, Z., Moloney, B., Wang, H.-T., … freec84. (2024). nipy/nibabel: 5.3.1 (5.3.1). Zenodo. https://doi.org/10.5281/zenodo.13936989

work page doi:10.5281/zenodo.13936989 2024

[6] [15]

Kingma, D. P. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

arXiv 2014

[7] [16]

R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D

Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., & Batra, D. (2017). Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision (pp. 618-626). Figures Figure 1. Proposed 4D CNN model architecture, consisting of four downsampling stages in a 1-1-...

work page 2017