Pith. sign in

REVIEW 4 major objections 6 minor 8 references

Age Sensitive Hippocampal Functional Connectivity: New Insights from 3D CNNs and Saliency Mapping

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Hippocampal connectivity lets a 3D CNN predict brain age within 7.3 years, and saliency maps reveal which connections matter most.

desk verdict A solid application of 3D CNN plus LayerCAM to hippocampal aging FC, with real novelty in the anterior/posterior split, though the saliency maps need validation before their specific regions are treated as findings. read the letter →

arxiv 2507.01411 v1 pith:AWZJTYRD submitted 2025-07-02 q-bio.NC cs.AIcs.CV

classification q-bio.NCcs.AIcs.CV
keywords brainagepredictionhippocampalfunctionalconnectivity3DconvolutionalneuralnetworkLayerCAMsaliencymappingseed-basedanterior-posteriorhippocampushealthyaging
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 tests whether the functional connections between the hippocampus and the rest of the brain contain enough information about healthy aging to predict a person's chronological age. Using voxel-wise hippocampal-cortical connectivity maps from 708 adults aged 36–90, the authors train a three-dimensional convolutional neural network (3D CNN) and report a validation mean absolute error of 7.30 years and a Spearman correlation of 0.79. To make the model interpretable, they adapt LayerCAM to produce voxel-level saliency maps, which highlight hippocampal connections with the precuneus, cuneus, posterior cingulate cortex, parahippocampal cortex, left superior parietal lobule, and right superior temporal sulcus as the most influential for age prediction. They further show that splitting the hippocampus into anterior and posterior parts gives comparable prediction accuracy but distinctly different saliency patterns, aligned with the known functional specializations of those subregions. If these results hold, they offer a fine-grained, interpretable view of how hippocampal connectivity reorganizes with age and a possible functional marker for age-related cognitive decline.

What carries the argument

The machinery is a combination of three pieces. First, seed-based functional connectivity: the time series of the left and right hippocampus (or their anterior/posterior halves) are averaged and correlated with every brain voxel, producing a pair of whole-brain 3D correlation volumes per subject. Second, a five-block 3D CNN with alternating 3×3×3 and 5×5×5 convolution kernels, batch normalization, ReLU, and max pooling, trained with mean absolute error as the loss for the regression task of predicting age. Third, LayerCAM, a class-activation-mapping variant that fuses positively weighted activation maps from all convolutional layers to yield high-resolution voxel-level saliency maps; the paper adapts it from 2D to 3D and from classification to regression. The saliency maps are the central interpretative device: they are averaged across subjects and used to identify which hippocampal–cortical connections are age-sensitive.

What would settle it

Replace the hippocampal FC values in the six highlighted regions with their mean (an occlusion test) and measure the change in predicted age; if the model's prediction is insensitive to those regions, the saliency maps do not reflect the model's true dependence. A second check would be to train the same pipeline on an independent aging sample and see whether the same cortical regions survive in the averaged saliency maps.

Watch

Extended reading notes

Core claim

The central claim is that seed-based hippocampal functional connectivity, when fed to a 3D CNN as a regression target, can predict chronological age with an error comparable to region-based connectivity models, and that the model's LayerCAM saliency maps identify specific hippocampal–cortical connections whose age sensitivity is not captured by linear regression. The whole-hippocampus model achieves a validation MAE of 7.30 years and Spearman rho 0.79 on the HCP-A cohort after pretraining on young adults and fivefold cross-validation fine-tuning. The averaged saliency map emphasizes the precuneus, cuneus, posterior cingulate cortex, parahippocampal cortex, left superior parietal lobule, and right superior temporal sulcus. The paper also claims that anterior and posterior hippocampal FC produce statistically indistinguishable prediction errors but significantly different saliency patterns, with anterior hippocampus more influential through prefrontal, anterior cingulate, entorhinal, and temporal pole connections, and posterior hippocampus more influential through precuneus, inferior parietal, and superior temporal connections. The authors interpret this as evidence that the model has captured functionally distinct aging trajectories along the hippocampal long axis.

Load-bearing premise

The paper's biological conclusions depend on the assumption that the subject-averaged LayerCAM saliency maps faithfully reflect which hippocampal–cortical connections actually drive aging, rather than dataset-specific confounds or model artifacts.

Editorial extensions

If this is right

  • The age signal in hippocampal–cortical FC is strong enough to predict chronological age within about 7.3 years, a performance comparable to region-based FC methods while preserving voxel-level spatial detail.
  • The six highlighted regions (precuneus, cuneus, PCC, PHC, left SPL, right STS) are candidate functional markers of healthy aging; their hippocampal FC declines with age and fits second- or third-order polynomial regressions better than linear ones.
  • Anterior and posterior hippocampal FC contribute to age prediction through different cortical targets, so analyzing the hippocampus as a single seed conceals subregion-specific aging effects.
  • The model is most accurate for 55–60-year-olds (MAE 4.69 years), suggesting the hippocampal FC age signal is strongest in middle age and less detectable in younger and older extremes.
  • Saliency maps are largely stable across sex and age groups, but subtle regional differences (e.g., precuneus and IPL more salient in females) hint that the model picks up sex-specific aging nuances without being trained on sex.

Reading between the lines

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

  • If the saliency maps truly reflect age-sensitive connectivity, the same pipeline could be applied to clinical cohorts (e.g., mild cognitive impairment or Alzheimer's disease) to see whether the highlighted hippocampal–cortical links are accelerated or redirected in pathological aging; this is an extension the paper does not test.
  • Because the reported 7.30-year MAE comes from cross-validation within a single dataset, an independent external cohort test would be a stronger check of generalization; the paper's comparisons to other studies use different datasets, so the numbers are not directly comparable.
  • The anterior/posterior dissociation suggests a testable prediction: if anterior hippocampus supports emotion and memory encoding and posterior supports spatial navigation, then the relative saliency of the two maps should predict which cognitive domain declines first in longitudinal data.
  • The nonlinear, spatially distributed features the CNN exploits (e.g., in precuneus and PCC, where linear regression finds almost nothing) could be extracted as a feature vector and correlated with known biomarkers of aging, such as hippocampal volume or amyloid burden, as a way to validate the saliency interpretation.
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.

Referee Report

4 major / 6 minor

Summary. The paper proposes an interpretable deep-learning pipeline for studying hippocampal aging. A 3D CNN is trained on voxel-wise seed-based functional connectivity maps from left and right hippocampal seeds, pretrained on HCP-Y and fine-tuned on HCP-A with fivefold cross-validation, achieving a validation MAE of 7.30 years and Spearman's rho of 0.79. The authors adapt LayerCAM to 3D regression to produce saliency maps, which they interpret as identifying hippocampal-cortical connections most influential for age prediction, specifically to precuneus, cuneus, PCC, PHC, left SPL, and right STS. They further compare whole-hippocampus, anterior-hippocampus, and posterior-hippocampus models, reporting similar prediction performance but distinct saliency patterns, which they interpret as reflecting known anterior/posterior functional specializations. Secondary analyses address sex differences, comparison with linear-regression age effects, and regional polynomial regressions.

Significance. If the saliency maps are faithful to the model's decision process, the paper offers a voxel-level, interpretable account of hippocampal-cortical connectivity changes in aging, and the anterior/posterior dissociation would be a useful contribution to the aging literature. The predictive performance is on par with or better than several region-based resting-state fMRI models, and the explicit adaptation of LayerCAM to 3D regression is a useful methodological step. The paper benefits from standard fivefold cross-validation, pretraining/fine-tuning, and comparison against several 3D CNN baselines. However, the central neurobiological claims rest entirely on saliency maps that are never validated as faithful explanations, and the performance claims lack fold-wise variance and may be affected by model selection on the same dataset. The significance is therefore conditional on additional validation experiments.

major comments (4)
  1. [Section 2.2 and Section 5.5] The central interpretability claim—that the highlighted regions are 'most influential' and 'age-sensitive'—rests on LayerCAM saliency maps (Eqs. 2-6) that are never validated as faithful explanations of the trained 3D CNN. No occlusion, perturbation, label-shuffling, or weight-randomization experiment is reported to show that these maps track the model's actual input-output dependence. The comparison with Grad-CAM (Suppl. Sec. 5) only shows higher spatial resolution, not correctness; the comparison with average FC and linear-regression z-scores (Figs. 2e-f) shows that saliency is different, not that it is more valid. Please add quantitative sanity checks, such as measuring the change in predicted age after occluding highlighted versus non-highlighted regions, or comparing saliency maps from a model trained on shuffled ages to rule out dataset-specific artifacts.
  2. [Section 5.5, Eqs. 2 and 4] The adaptation of LayerCAM to regression keeps only positive gradients (Eq. 2) and applies a ReLU to the combined map (Eq. 4), thereby discarding all negative-age-direction contributions. For a regression task, a voxel whose hippocampal FC decreases with age can produce negative gradients, and truncating them biases the saliency map toward positive-direction effects. The claim that the highlighted regions are 'most influential' is therefore incomplete; please justify the sign truncation or compare the reported maps with full-gradient or integrated-gradient alternatives to show that the discarded information is nonessential.
  3. [Section 2.4 and Section 5.8] The anterior versus posterior saliency comparison (Fig. 5c) does not control for the different baseline FC distributions of the two seed regions. The anterior and posterior hippocampus have known anatomical connectivity differences (ref. [18]), so the observed saliency differences may simply mirror these baseline connectivity differences rather than age-specific sensitivity. A control analysis, such as matching the FC-strength distributions or comparing the saliency differences against a null distribution obtained from label-shuffled models, is needed before interpreting Fig. 5c as evidence of 'distinct functional trajectories in normal aging.'
  4. [Section 2.1 and Suppl. Sec. 3] The reported validation MAE of 7.30 years may be optimistically biased because the architecture was selected on the same HCP-A dataset via fivefold cross-validation (Suppl. Table 1), and no fold-wise variance or confidence intervals are reported. Please provide per-fold MAE values and either use nested cross-validation or clearly state the model-selection procedure so readers can gauge the selection-induced optimism.
minor comments (6)
  1. [Abstract] The phrase 'remain s limited' in the abstract contains a typo and should read 'remains limited.'
  2. [Section 5.6] The text says 'excluding subjects aged 100-year-old,' which is inconsistent with Section 5.1's description of excluding 12 subjects aged above 90 because detailed age information was not provided; please reconcile the age cutoff.
  3. [Fig. 3 caption] The caption states that 'MAE values (in years) were denoted on the right side of each bar,' but these values are not visible in the text version; please ensure the figure actually displays the numerical MAE values.
  4. [Sections 2.4, 5.7, 5.8] The phrase 'one-way paired t-tests' should be 'one-sided paired t-tests' or 'one-tailed paired t-tests' to correctly describe the statistical test used.
  5. [Suppl. Sec. 6] The text mentions 'left superior temporal sulcus (right STS)' and 'right superior parietal lobule (right SPL),' which contradict the main-text naming of 'left SPL and right STS'; please correct these labels.
  6. [Suppl. Sec. 6] Supplementary table numbering is inconsistent (a 'Suppl. Table1' is referenced for model selection in Sec. 3, and the regression results are labeled both 'Suppl. Table1' and 'Suppl. Table2'); please harmonize all supplementary table and figure references.

Circularity Check

1 steps flagged · score 6.0 of 10

Central saliency-based claim reduces to the definition of LayerCAM: the 'age-sensitive regions' are, by construction, the regions that contributed most to the fitted age-prediction model.

  1. self definitional [Sec. 2.2, Sec. 5.5 (Eqs. 2-6), Sec. 3.1]
    "In the saliency map, brighter regions in the saliency map indicate a greater influence on the model prediction. ... the saliency maps generated from LayerCAM would highlight the regions that contributed the most to making the specific age prediction as a continuous number. ... Based on the averaged saliency map, the hippocampal FC model emphasized connections between hippocampus and areas including PCC, PHC, left SPL, right STS, and precuneus ... By generating saliency maps, we identified brain regions whose hippocampal FC is highly relevant to the normal aging process."

    Section 5.5 defines the saliency map M via Eqs. 2-4 as a ReLU combination of gradient-weighted feature maps, using the positive gradients of the age-prediction output; the text explicitly says these maps 'highlight the regions that contributed the most to making the specific age prediction.' Section 2.2 then presents those same maps as evidence about 'regions ... highly sensitive to age' and the conclusion states the model 'identified brain regions whose hippocampal FC is highly relevant to the normal aging process.' Because saliency is constructed as contribution-to-age-prediction, the statement that the highlighted regions are the ones most influential for age prediction is true by definition; it is a description of the fitted CNN, not an independent empirical derivation.

full rationale

The paper's age-prediction results are not circular: the 3D CNN is pretrained on HCP-Y and fine-tuned with fivefold cross-validation on HCP-A, and the reported validation MAE (7.30 years) and Spearman rho (0.79) are genuine held-out quantities compared against external benchmarks. The circularity is confined to the interpretability layer. In Section 5.5, the adapted LayerCAM saliency M is defined (Eqs. 2-4) from gradients of the age-prediction output, and the text states that it highlights regions that contributed most to the age prediction. Section 2.2 and the Discussion then convert this self-referential quantity into a biological finding: 'regions whose hippocampal FC is highly relevant to the normal aging process.' Since the saliency map is by construction an attribution for the fitted age-prediction task, saying that salient connections are 'age-sensitive' or 'most influential for age prediction' is a definitional restatement rather than an independent result. The paper's checks (overlap with average FC, comparison with linear-regression z-scores, Grad-CAM resolution comparison) show only that LayerCAM differs from those maps; they do not validate the maps against ground-truth model reliance or external aging markers. The anterior/posterior saliency comparison (Sec. 2.4) has the same structure: differences between models trained on anterior vs posterior seeds are model attributions, then interpreted as 'functional segregation' of aging. The self-citation [18] for anterior-posterior anatomical connectivity is not load-bearing here because the functional dissociation is also supported by external references [19-22,55,57]. Overall, the predictive core is externally grounded, but the paper's headline interpretability claim is partly circular by construction, giving a score of 6.

Assumptions & free parameters 9 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new entities. The model's hyperparameters are chosen by hand and are listed as free parameters. The central assumptions are neuroimaging domain assumptions about FC reliability, cohort transferability, manual segmentation, and the validity of LayerCAM attributions.

free parameters (9)
  • Learning rate = 0.00002
    Chosen for pretraining and fine-tuning (Section 5.4); affects convergence and final performance.
  • Batch size = 16
    Chosen in Section 5.4; affects optimization.
  • Convolutional filter counts = 16, 16, 32, 32, 64
    Architecture design in Section 5.3; chosen by hand, not optimized.
  • Kernel sizes = 3x3x3 and 5x5x5 alternating
    Architecture choice motivated by Inception modules (Section 5.3).
  • Max-pooling kernel and stride = 2x2x2, stride 2
    Architecture design in Section 5.3.
  • LayerCAM scaling factor gamma = 2
    Taken from the original LayerCAM paper (Section 5.5, Eq. 5); not optimized for this task.
  • Pretraining epochs = 50
    Set in Section 5.4; fixed for pretraining.
  • Early stopping patience = 15
    Set in Section 5.4 for fine-tuning.
  • Maximum fine-tuning epochs = 200
    Set in Section 5.4 for fine-tuning.
assumptions (5)
  • domain assumption Resting-state fMRI data from HCP-Y and HCP-A are comparable enough for pretraining and fine-tuning after minimal preprocessing, smoothing, and bandpass filtering.
    The model is pretrained on HCP-Y (22-37 years) and fine-tuned on HCP-A (36-90 years); this transfer assumes the FC-age relationship is stable across cohorts and protocols (Sections 5.1, 5.4).
  • domain assumption Pearson correlation between averaged hippocampal seed time series and every cortical voxel, averaged over four fMRI runs, yields a reliable estimate of hippocampal functional connectivity.
    The entire input representation relies on this standard seed-based FC computation (Section 5.2).
  • domain assumption The manual split of the hippocampus into anterior and posterior parts by an experienced neuropsychologist is sufficiently accurate for all subjects.
    Anterior and posterior FC models depend on this segmentation; the authors report only 'minor misalignment' after reverse registration (Section 5.8).
  • domain assumption LayerCAM gradients, originally designed for classification, provide a meaningful attribution of the model's regression output to input voxels.
    The paper adapts LayerCAM to regression and assumes the saliency maps represent contribution to the continuous age prediction (Section 5.5).
  • domain assumption The MMP atlas parcellation used to label saliency regions matches the functional organization relevant to aging.
    Region names for saliency findings are assigned via the MMP atlas overlay (Figure 2, Section 5.6).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Age Sensitive Hippocampal Functional Connectivity: New Insights from 3D CNNs and Saliency Mapping." pith.science (2026). https://pith.science/paper/AWZJTYRD

@misc{pith2026250701411,
  author       = {Pith},
  title        = {Pith review of: Age Sensitive Hippocampal Functional Connectivity: New Insights from 3D CNNs and Saliency Mapping},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AWZJTYRD}},
  note         = {Machine review of arXiv:2507.01411}
}
read the original abstract

Grey matter loss in the hippocampus is a hallmark of neurobiological aging, yet understanding the corresponding changes in its functional connectivity remains limited. Seed-based functional connectivity (FC) analysis enables voxel-wise mapping of the hippocampus's synchronous activity with cortical regions, offering a window into functional reorganization during aging. In this study, we develop an interpretable deep learning framework to predict brain age from hippocampal FC using a three-dimensional convolutional neural network (3D CNN) combined with LayerCAM saliency mapping. This approach maps key hippocampal-cortical connections, particularly with the precuneus, cuneus, posterior cingulate cortex, parahippocampal cortex, left superior parietal lobule, and right superior temporal sulcus, that are highly sensitive to age. Critically, disaggregating anterior and posterior hippocampal FC reveals distinct mapping aligned with their known functional specializations. These findings provide new insights into the functional mechanisms of hippocampal aging and demonstrate the power of explainable deep learning to uncover biologically meaningful patterns in neuroimaging data.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

8 extracted references · 5 canonical work pages

  1. [1]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, "Very deep convolutional networks for large-scale image recognition," arXiv preprint arXiv:1409.1556, 2014

  2. [2]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image recognition," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770-778

  3. [3]

    Efficientnet: Rethinking model scaling for convolutional neural networks,

    M. Tan and Q. Le, "Efficientnet: Rethinking model scaling for convolutional neural networks," in International conference on machine learning, 2019: PMLR, pp. 6105-6114

  4. [4]

    Accurate brain age prediction with lightweight deep neural networks,

    H. Peng, W. Gong, C. F. Beckmann, A. Vedaldi, and S. M. Smith, "Accurate brain age prediction with lightweight deep neural networks," Medical image analysis, vol. 68, p. 101871, 2021

  5. [5]

    Deep learning-based brain age prediction in normal aging and dementia,

    J. Lee et al., "Deep learning-based brain age prediction in normal aging and dementia," Nature Aging, vol. 2, no. 5, pp. 412-424, 2022

  6. [6]

    The (limited?) utility of brain age as a biomarker for capturing cognitive decline,

    A. Tetereva and N. Pat, "The (limited?) utility of brain age as a biomarker for capturing cognitive decline," eLife, vol. 12, 2023

  7. [7]

    Layercam: Exploring hierarchical class activation maps for localization,

    P.-T. Jiang, C.-B. Zhang, Q. Hou, M.-M. Cheng, and Y. Wei, "Layercam: Exploring hierarchical class activation maps for localization," IEEE Transactions on Image Processing, vol. 30, pp. 5875-5888, 2021

  8. [8]

    Differential effect of age on posterior and anterior hippocampal functional connectivity,

    J. S. Damoiseaux, R. P. Viviano, P. Yuan, and N. Raz, "Differential effect of age on posterior and anterior hippocampal functional connectivity," Neuroimage, vol. 133, pp. 468-476, 2016

Pith tools

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