Pith. sign in

REVIEW 3 major objections 7 minor 67 references

Hyperbolic Kernel Graph Neural Networks for Neurocognitive Decline Analysis from Multimodal Brain Imaging

T0 review · 3 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Hyperbolic kernel graph networks outperform Euclidean baselines on early neurocognitive decline classification.

desk verdict Solid empirical fusion framework; the hyperbolic geometry claim is not isolated and likely not what drives the gains. read the letter →

arxiv 2507.02908 v1 pith:TX6QDREG submitted 2025-06-24 cs.LG cs.AI

classification cs.LGcs.AI
keywords hyperbolickernelgraphneuralnetworksmultimodalneuroimagingbrainconnectivityneurocognitivedeclineDTIandfMRIfusionPoincaréballclassificationtransferlearning
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 claims that brain connectivity graphs—structural networks from diffusion MRI (DTI) and functional networks from resting-state fMRI—are hierarchically organized, and that Euclidean graph neural networks miss this organization. It introduces HKGF, a framework that encodes these graphs with hyperbolic kernel graph neural networks (HKGNNs), fuses structural and functional embeddings through a learned SC-FC coupling graph, and classifies with a two-layer hyperbolic network. On two small clinical cohorts, the attention-based variant HKGF2 reaches AUC 80.42% on SMC vs CN classification in ADNI and 69.74% on ANI vs CN classification in HAND, outperforming Euclidean GCN/GAT variants and a prior hyperbolic GCN. The paper also shows the framework transfers to fMRI+ASL inputs and benefits from pretraining on 3,806 auxiliary fMRI scans.

What carries the argument

The load-bearing object is the hyperbolic kernel graph layer, instantiated as HKGCN (Eq. (14)) and HKGAT (Eq. (18)). Both versions project node features into the Poincaré ball, apply the logarithmic map to reach the tangent space, perform adjacency- or attention-based aggregation there, and then apply a two-term activation: a nonlinearity $f(\cdot)$ plus a scaled cosine $\lambda\cos(\cdot)$. The first term is a topology-aware approximation of the hyperbolic arc-cos (HAC) kernel and is meant to capture global relationships; the cosine term approximates the hyperbolic RBF (HRBF) kernel and captures local similarity. The SC-FC coupling graph, built as the inner product of normalized structural and functional embeddings, is the mechanism that fuses modalities before a second HKGNN pass. This design avoids Möbius addition and multiplication, which is why the paper reports HKGCN matching GCN in FLOPs and HKGAT matching GAT.

What would settle it

Train HKGF alongside a matched Euclidean control that replaces the logarithmic map with the identity (equivalently, sets curvature $c$ to 0) while keeping the same ReLU-plus-cosine activations and the same coupling graph; if ADNI SMC-vs-CN AUC stays at the reported 80.42%, then the hyperbolic geometry is not what produces the improvement.

Watch

Extended reading notes

Core claim

The paper's central discovery is that replacing Euclidean message passing with two curvature-aware kernels—a hyperbolic arc-cos kernel for global similarity and a hyperbolic RBF kernel for local similarity—lets a graph neural network represent brain networks in a Poincaré ball at roughly the computational cost of a standard GCN. Node features are projected into the ball, pulled back to the tangent space with the logarithmic map $\log_0^c$, aggregated over the graph, and then activated by $f(\cdot)+\lambda\cos(\cdot)$, where the cosine term approximates an HRBF kernel. The same kernel layers are applied a second time to a data-driven coupling graph whose edges are the inner product of normalized structural and functional embeddings, and a two-layer hyperbolic network performs prediction. The paper reports consistent gains over Euclidean baselines and over HGCN in both tasks, with HKGF2 (the attention backbone) giving the best numbers.

Load-bearing premise

The framework's load-bearing premise is that mapping brain-network nodes into a Poincaré ball and computing kernels in the tangent space captures the hierarchical organization that Euclidean graph networks miss; the paper does not directly measure whether the learned embeddings are actually more hierarchical or tree-like.

Editorial extensions

If this is right

  • If correct, DTI-fMRI fusion for early cognitive decline can run in hyperbolic space at essentially the same FLOPs as Euclidean GCN/GAT, removing the main practical barrier of prior hyperbolic GNNs.
  • The framework is modality-agnostic: the authors show fMRI+ASL classification reaches AUC 76.85%, so the same coupling and kernel machinery can absorb perfusion imaging alongside connectivity data.
  • Pretraining on large auxiliary fMRI cohorts adds more than 5% AUC on the HAND task, suggesting transfer learning can offset the tiny target cohorts typical of clinical neuroimaging studies.
  • The discriminative regions the method highlights—frontal-parietal ROIs for SMC, subcortical and cerebellar ROIs for ANI—are consistent with prior literature, so the model can point toward plausible biomarkers rather than only a scalar risk score.

Reading between the lines

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

  • The paper does not measure tree-likeness or hyperbolicity of the learned embeddings, so a direct test of the mechanism would compare HKGF against its Euclidean counterpart with identical two-term activations and coupling; if the gap collapses, the performance gain is not coming from negative curvature.
  • The sensitivity analysis shows curvature $c$ has only a minor effect while the cosine weight $\lambda$ matters, which suggests the practical boost may come from the residual cosine branch and the coupling graph rather than from hyperbolic geometry per se.
  • The inner-product coupling graph could be reused for other heterogeneous graph pairs, such as gene co-expression with protein interaction or behavioral measures with connectivity, giving a general fusion primitive beyond neuroimaging.
  • Because HKGF requires complete modality pairs and the authors flag missing-modality handling as future work, a clinical deployment would need imputation or a partial-observation extension before the framework can be used on incomplete records.
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

3 major / 7 minor

Summary. The paper introduces HKGF, a multimodal fusion framework for neurocognitive decline classification that combines DTI- and fMRI-derived brain graphs (and, in a generalization experiment, ASL-derived graphs) using a family of hyperbolic kernel graph neural networks (HKGCN and HKGAT). The HKGNNs project node features into a Poincaré ball and apply logarithmic maps to a tangent space before performing graph aggregation with ReLU/ELU and cosine activations; a cross-modality coupling graph and a hyperbolic neural network are used for fusion and prediction. The authors report state-of-the-art results on SMC vs. CN classification on ADNI (fMRI+DTI) and ANI vs. CN classification on HAND, with a transfer-learning strategy pretraining models on 3,806 auxiliary fMRI scans. The paper includes ablations, hyperparameter sensitivity, generalization to ASL, computational complexity analysis, and visualization of discriminative brain regions.

Significance. If the reported results hold, HKGF would be a practically useful multimodal fusion framework for challenging neurocognitive decline tasks, with the notable strengths of released source code and pretrained models, extensive baselines, a large auxiliary pretraining corpus, and a computational-cost analysis showing parity with GCN/GAT. The central scientific claim, however, is that hyperbolic geometry is what drives the gains by preserving hierarchical brain-network structure; this claim is currently not established, because the hyperbolic component is not isolated and the paper's own sensitivity analysis weakens it. The contribution would be more credible as an empirical fusion method than as a demonstration of hyperbolic representation benefits.

major comments (3)
  1. [§5.2, Eqs. (14) and (18)] The hyperbolic component is not isolated, and the paper's own evidence suggests it is not the driver of performance. With the implemented curvature c=0.001 and the projection in Eq. (13) leaving most features unchanged, the logarithmic map in Eq. (4) satisfies log_c^0(z) ≈ z + O(c), so Eqs. (14) and (18) are effectively Euclidean GCN/GAT layers with an extra λ-scaled cosine branch. Figure 5 confirms that the curvature parameter c has only a minor effect on AUC/ACC, while λ has a more noticeable impact. The ablations HKGF-G and HKGF-A replace HKGCN/HKGAT with plain GCN/GAT, changing both the log map and the cosine branch simultaneously, so they cannot attribute the improvement to hyperbolic geometry. The paper also reports no δ-hyperbolicity, tree-likeness, or hierarchy-preservation metric for the learned embeddings. To support the central claim, the authors should compare against an identical architecture with the log map replaced by the identity (or a large c range), and measure whether the embeddings actually reflect hierarchical structure.
  2. [§3.2.2, Eqs. (7)–(12) vs. Eq. (14)] The kernel derivation is not faithfully implemented. The HAC and HRBF kernels are defined via random feature expansions with fixed weights W drawn from p(w), as in Eqs. (8) and (11), but the HKGCN and HKGAT layers in Eqs. (14) and (18) use trainable weights and apply nonlinear activations after graph aggregation, and the cosine term is applied to the aggregated features rather than to pointwise random Fourier features. No analysis or theorem is provided to show that Eq. (14) approximates the proposed hyperbolic kernels on graphs. Since the method is named and motivated by hyperbolic kernel theory, this gap between the theoretical formulation and the implemented architecture should be addressed or the kernel framing should be softened.
  3. [§4.2.2 vs. Tables 2 and 6] There is a direct contradiction in the experimental setup. Section 4.2.2 defines Task 1 as SMC vs. CN classification on ADNI with fMRI and ASL data from 29 SMC and 15 CN subjects, but Table 2 reports SMC vs. CN results on ADNI with fMRI and DTI data from 46 SMC and 48 CN subjects, and the ASL experiment appears only later in Table 6. This inconsistency makes it unclear what data underpin the main ADNI claim and must be corrected for reproducibility.
minor comments (7)
  1. [§4.2.2] There is a typo in the first sentence: "Twp prediction tasks" should be "Two prediction tasks."
  2. [Eq. (14)] The notation b^T in f(Â(log_c^0(X̃)W + b^T)) is confusing: since b ∈ R^M is defined as a vector, adding a transposed vector to an N×M matrix is not well-defined. Please clarify whether this is meant to be a broadcast bias term, and use consistent notation (e.g., 1b^T).
  3. [Fig. 5 caption] The caption refers to "different settings of c and α," but the hyperparameter discussed in the text and figures is λ. Please correct the symbol.
  4. [§5.1] In the description of ablations, "HKGFC-A" appears to be a typo for "HKGF-A."
  5. [§5.4] The word "prepossessing" in the ASL generalization section should be "preprocessing."
  6. [§2.1, reference [22]] The text attributes a dynamic-FC GNN framework to "Liu et al. [22]," but reference [22] is Ereira et al., Nature Mental Health 2024, which appears to be a different work. Please verify the citation or replace it with the intended reference.
  7. [Tables 2 and 4] The reported HKGF2 results differ slightly between Tables 2 and 4 (e.g., AUC 80.42 vs. 80.36, ACC 81.26 vs. 81.83). Please explain whether these are different random seeds or a reporting error.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; minor self-citations for the HAND dataset and ACTION pretraining toolbox are not load-bearing.

full rationale

I find no circular step in the claimed derivation chain. The central prediction is classification performance on ADNI and HAND, which is measured on held-out folds against many external baselines; the architecture is not fitted to or derived from the test labels. The hyperbolic kernels in Eqs. (7)-(12) are explicitly defined as Euclidean arc-cos/RBF kernels applied to log_c^0(z), so they are definitions rather than hidden reuses of the outcome. The HKGCN/HKGAT layers in Eqs. (14) and (18) are new parameterized models, not the kernel values themselves; the kernel-to-architecture link is informal, but that is a methodological weakness rather than circularity. The small curvature c=0.001 makes the log map nearly identity for the given features, weakening the claim that hyperbolic geometry drives the gains; this is a correctness/ablation concern, not a circular reduction. The only self-citations are [23] for the HAND dataset source and [56] for the ACTION pretraining toolbox; both serve as data/code provenance, and the pretraining contribution is directly evaluated via the w/oP ablation in Table 5. No load-bearing self-citation, uniqueness theorem, or fitted-input-renamed-as-prediction pattern was identified, so the paper is essentially self-contained against external benchmarks.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests mainly on the assumption that tangent-space operations after a logarithmic map preserve hierarchical brain structure, plus a handful of hand-set hyperparameters. There are no new physical entities. The most fragile part is the unmeasured link between hyperbolic geometry and the reported classification gains.

free parameters (4)
  • curvature c = 0.001
    Hand-set value defining the Poincaré ball geometry. Sensitivity analysis in Fig. 5 shows it has a minor effect on performance, which weakens the claim that hyperbolic geometry drives the gains.
  • kernel balance lambda = 0.01
    Hand-set scalar balancing the ReLU branch and cosine branch in HKGCN and HKGAT. Fig. 5 shows performance varies more with lambda than with curvature, so this tuning choice matters.
  • FC graph edge threshold = top 50% strongest edges
    Empirical choice following reference [39] to sparsify fully connected functional connectivity graphs. This changes graph topology and therefore the learned features.
  • random feature dimension M = not explicitly reported; hidden dim 64
    The kernel approximation quality depends on the number of random features, but the paper does not state how many random vectors are sampled for HAC and HRBF.
assumptions (5)
  • domain assumption Brain structural and functional networks have a hierarchical organization that hyperbolic geometry can represent.
    Motivates the entire framework in Section 3 and Fig. 2. The cited literature supports hierarchy, but the paper does not measure tree-likeness or hyperbolic structure of its learned embeddings.
  • ad hoc to paper The logarithmic map to the tangent space, followed by Euclidean operations, preserves the hierarchical information of the Poincaré ball.
    Used in Eq. (14), Eq. (15), and Eq. (21). All trainable operations happen in tangent space, and the paper provides no formal guarantee that hierarchy is preserved.
  • standard math Random feature approximation of the HAC and HRBF kernels is valid in the graph aggregation setting.
    Borrowed from Rahimi and Recht [41] and Cho and Saul [40]. The approximation quality inside a graph neural network is not analyzed.
  • standard math Tangent-space Euclidean distance approximates hyperbolic distance with a scaling factor, per the Curve Length Equivalence Theorem in [36].
    Used to justify HRBF as a local hyperbolic similarity in Eq. (9) and Eq. (10). The paper relies on this cited theorem without re-deriving it.
  • domain assumption The AAL atlas with 116 ROIs defines meaningful and comparable nodes across both structural and functional modalities.
    All graphs use these predefined regions, and the entire pipeline inherits any errors or limitations of the atlas alignment across subjects and modalities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hyperbolic Kernel Graph Neural Networks for Neurocognitive Decline Analysis from Multimodal Brain Imaging." pith.science (2026). https://pith.science/paper/TX6QDREG

@misc{pith2026250702908,
  author       = {Pith},
  title        = {Pith review of: Hyperbolic Kernel Graph Neural Networks for Neurocognitive Decline Analysis from Multimodal Brain Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TX6QDREG}},
  note         = {Machine review of arXiv:2507.02908}
}
read the original abstract

Multimodal neuroimages, such as diffusion tensor imaging (DTI) and resting-state functional MRI (fMRI), offer complementary perspectives on brain activities by capturing structural or functional interactions among brain regions. While existing studies suggest that fusing these multimodal data helps detect abnormal brain activity caused by neurocognitive decline, they are generally implemented in Euclidean space and can't effectively capture intrinsic hierarchical organization of structural/functional brain networks. This paper presents a hyperbolic kernel graph fusion (HKGF) framework for neurocognitive decline analysis with multimodal neuroimages. It consists of a multimodal graph construction module, a graph representation learning module that encodes brain graphs in hyperbolic space through a family of hyperbolic kernel graph neural networks (HKGNNs), a cross-modality coupling module that enables effective multimodal data fusion, and a hyperbolic neural network for downstream predictions. Notably, HKGNNs represent graphs in hyperbolic space to capture both local and global dependencies among brain regions while preserving the hierarchical structure of brain networks. Extensive experiments involving over 4,000 subjects with DTI and/or fMRI data suggest the superiority of HKGF over state-of-the-art methods in two neurocognitive decline prediction tasks. HKGF is a general framework for multimodal data analysis, facilitating objective quantification of structural/functional brain connectivity changes associated with neurocognitive decline.

Figures

Figures reproduced from arXiv: 2507.02908 by the authors.

Figure 1
Figure 1. Illustration of the proposed hyperbolic kernel graph fusion (HKGF) framework for neurocognitive decline analysis with multimodal data. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. T-SNE [30] visualizations of regional (a) fMRI and (b) DTI features for subjects from HAND cohort [23]. For fMRI data, each brain ROI defined by AAL atlas is represented by a 116-dimensional vector, where each element corresponds to the functional connectivity (measured using Pearson correlation coefficients) with all other ROIs. For DTI data, each ROI is represented by a 348-dimensional feature vector, capturing it… view at source ↗
Figure 3
Figure 3. The t-SNE [30] visualization of features output by (a) HKGCN backbone and (b) HKGAT backbone in single-modality graph representation learning (1st stage) and cross-modality coupling (2nd stage) modules of the proposed HKGF1 and HKGF2, respectively. Shown for (top) Task 1: SMC vs. CN classification on the ADNI dataset and (bottom) ANI vs. CN classification on the HAND dataset using fMRI and DTI data. Axial View Coron… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Discriminative brain regions identified by HKGF [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Influence of the two hyperparameters (c and λ) on the performance of (a) HKGF1 and (b) HKGF2 for ANI vs. CN classification on HAND. As reported in [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

67 extracted references · 65 canonical work pages

  1. [1]

    Complementary aspects of diffusion imaging and fMRI; I: structure and function,

    R. V . Mulkern, P . E. Davis, S. J. Haker, R. S. J. Estepar, L. P . Panych, S. E. Maier, and M. J. Rivkin, “Complementary aspects of diffusion imaging and fMRI; I: structure and function,”Magnetic Resonance Imaging, vol. 24, no. 4, pp. 463–474, 2006

  2. [2]

    MR diffusion tensor spectroscopy and imaging,

    P . J. Basser, J. Mattiello, and D. LeBihan, “MR diffusion tensor spectroscopy and imaging,”Biophysical Journal, vol. 66, no. 1, pp. 259–267, 1994

  3. [3]

    Functional connectivity in the motor cortex of resting human brain using echo-planar MRI,

    B. Biswal, F. Zerrin Yetkin, V . M. Haughton, and J. S. Hyde, “Functional connectivity in the motor cortex of resting human brain using echo-planar MRI,”Magnetic Resonance in Medicine, vol. 34, no. 4, pp. 537–541, 1995

  4. [4]

    Spontaneous fluctuations in brain activity observed with functional magnetic resonance imaging,

    M. D. Fox and M. E. Raichle, “Spontaneous fluctuations in brain activity observed with functional magnetic resonance imaging,” Nature Reviews Neuroscience, vol. 8, no. 9, pp. 700–711, 2007

  5. [5]

    Perfusion magnetic resonance imag- ing with continuous arterial spin labeling: Methods and clinical applications in the central nervous system,

    J. A. Detre and D. C. Alsop, “Perfusion magnetic resonance imag- ing with continuous arterial spin labeling: Methods and clinical applications in the central nervous system,”European Journal of Radiology, vol. 30, no. 2, pp. 115–124, 1999

  6. [6]

    D. C. Alsop, J. A. Detre, X. Golay, M. G ¨unther, J. Hendrikse, L. Hernandez-Garcia, H. Lu, B. J. MacIntosh, L. M. Parkes, M. Smitset al., “Recommended implementation of arterial spin- labeled perfusion MRI for clinical applications: A consensus of the ISMRM perfusion study group and the European consortium for ASL in dementia,”Magnetic Resonance in Medic...

  7. [7]

    Fusing DTI and fMRI data: A survey of methods and applications,

    D. Zhu, T. Zhang, X. Jiang, X. Hu, H. Chen, N. Yang, J. Lv, J. Han, L. Guo, and T. Liu, “Fusing DTI and fMRI data: A survey of methods and applications,”NeuroImage, vol. 102, pp. 184–191, 2014

  8. [8]

    Divergent network connectivity changes in behavioural variant frontotemporal dementia and Alzheimer’s disease,

    J. Zhou, M. D. Greicius, E. D. Gennatas, M. E. Growdon, J. Y. Jang, G. D. Rabinovici, J. H. Kramer, M. Weiner, B. L. Miller, and W. W. Seeley, “Divergent network connectivity changes in behavioural variant frontotemporal dementia and Alzheimer’s disease,”Brain, vol. 133, no. 5, pp. 1352–1367, 2010

Show all 67 references
  1. [9]

    The organization of the human cerebral cortex estimated by intrinsic functional connectivity,

    B. T. T. Yeo, F. M. Krienen, J. Sepulcre, M. R. Sabuncu, D. Lashkari, M. Hollinshead, J. L. Roffman, J. W. Smoller, L. Z ¨ollei, J. R. Polimeni, B. Fischl, H. Liu, and R. L. Buckner, “The organization of the human cerebral cortex estimated by intrinsic functional connectivity,...

  2. [10]

    Situating the default- mode network along a principal gradient of macroscale cortical organization,

    D. S. Margulies, S. Ghosh, A. Goulas, M. Falkiewicz, J. M. Huntenburg, G. Langs, G. Bezgin, S. B. Eickhoff, F. X. Castellanos, M. Petrides, E. Jefferies, and J. Smallwood, “Situating the default- mode network along a principal gradient of macroscale cortical organization,”Proc...

  3. [11]

    Hierarchical modularity in human brain functional networks,

    D. Meunier, R. Lambiotte, A. Fornito, K. Ersche, and E. Bullmore, “Hierarchical modularity in human brain functional networks,” Frontiers in Neuroinformatics, vol. 3, p. 37, 2009

  4. [12]

    Changes in structural and functional connectivity among resting- state networks across the human lifespan,

    R. F. Betzel, L. Byrge, Y. He, J. Goni, X. N. Zuo, and O. Sporns, “Changes in structural and functional connectivity among resting- state networks across the human lifespan,”NeuroImage, vol. 102, pp. 345–357, 2014

  5. [13]

    Hierarchical organization unveiled by functional connectivity in complex brain networks,

    C. Zhou, L. Zemanov ´a, G. Zamora, C. C. Hilgetag, and J. Kurths, “Hierarchical organization unveiled by functional connectivity in complex brain networks,”Physical Review Letters, vol. 97, no. 23, p. 238103, 2006

  6. [14]

    Re- vealing modular architecture of human brain structural networks by using cortical thickness from MRI,

    Z. J. Chen, Y. He, P . Rosa-Neto, J. Germann, and A. C. Evans, “Re- vealing modular architecture of human brain structural networks by using cortical thickness from MRI,”Cerebral Cortex, vol. 18, no. 10, pp. 2374–2381, 2008

  7. [15]

    ‘Hierarchy’ in the organization of brain networks,

    C. C. Hilgetag and A. Goulas, “‘Hierarchy’ in the organization of brain networks,”Philosophical Transactions of the Royal Society B, vol. 375, no. 1796, p. 20190319, 2020

  8. [16]

    Hyperbolic graph convolutional neural networks,

    I. Chami, R. Ying, C. R ´e, and J. Leskovec, “Hyperbolic graph convolutional neural networks,” inNeurIPS, vol. 32, 2019

  9. [17]

    Hyperbolic space sparse coding with its application on prediction of alzheimer’s disease in mild cognitive impairment,

    J. Zhang, J. Shi, C. Stonnington, Q. Li, B. A. Gutman, K. Chen, E. M. Reiman, R. Caselli, P . M. Thompson, J. Yeet al., “Hyperbolic space sparse coding with its application on prediction of alzheimer’s disease in mild cognitive impairment,” inMICCAI. Springer, 2016, pp. 326–334

  10. [18]

    Skin lesion recognition with class-hierarchy regularized hyperbolic embeddings,

    Z. Yu, T. Nguyen, Y. Gal, L. Ju, S. S. Chandra, L. Zhang, P . Bon- nington, V . Mar, Z. Wang, and Z. Ge, “Skin lesion recognition with class-hierarchy regularized hyperbolic embeddings,” inMICCAI. Springer, 2022, pp. 594–603

  11. [19]

    Multimodal deep fusion in hyperbolic space for mild cognitive impairment study,

    L. Zhang, S. Na, T. Liu, D. Zhu, and J. Huang, “Multimodal deep fusion in hyperbolic space for mild cognitive impairment study,” inMICCAI. Springer, 2023, pp. 674–684

  12. [20]

    DTI analyses and clinical applications in Alzheimer’s disease,

    K. Oishi, M. M. Mielke, M. Albert, C. G. Lyketsos, and S. Mori, “DTI analyses and clinical applications in Alzheimer’s disease,” Journal of Alzheimer’s Disease, vol. 26, no. s3, pp. 287–296, 2011

  13. [21]

    Cerebral abnormalities in HIV-infected individuals with neurocognitive impairment revealed by fMRI,

    P .-p. Chen, X.-y. Wei, L. Tao, X. Xin, S.-t. Xiao, and N. He, “Cerebral abnormalities in HIV-infected individuals with neurocognitive impairment revealed by fMRI,”Scientific Reports, vol. 13, no. 1, p. 10331, 2023

  14. [22]

    Early detection of dementia with default-mode network effective connectivity,

    S. Ereira, S. Waters, A. Razi, and C. R. Marshall, “Early detection of dementia with default-mode network effective connectivity,” Nature Mental Health, vol. 2, no. 7, pp. 787–800, 2024

  15. [23]

    Leveraging brain modularity prior for interpretable representation learning of fMRI,

    Q. Wang, W. Wang, Y. Fang, P .-T. Yap, H. Zhu, H.-J. Li, L. Qiao, and M. Liu, “Leveraging brain modularity prior for interpretable representation learning of fMRI,”IEEE Transactions on Biomedical Engineering, vol. 71, no. 8, pp. 2391–2401, 2024

  16. [24]

    Functional MRI-guided probabilistic tractography of cortico- cortical and cortico-subcortical language networks in children,

    P . J. Broser, S. Groeschel, T.-K. Hauser, K. Lidzba, and M. Wilke, “Functional MRI-guided probabilistic tractography of cortico- cortical and cortico-subcortical language networks in children,” NeuroImage, vol. 63, no. 3, pp. 1561–1570, 2012

  17. [25]

    Inferring functional connectivity in MRI using Bayesian network structure learning with a modified PC algorithm,

    S. P . Iyer, I. Shafran, D. Grayson, K. Gates, J. T. Nigg, and D. A. Fair, “Inferring functional connectivity in MRI using Bayesian network structure learning with a modified PC algorithm,”NeuroImage, vol. 75, pp. 165–175, 2013

  18. [26]

    Multi- modal graph neural network for early diagnosis of Alzheimer’s disease from sMRI and PET scans,

    Y. Zhang, X. He, Y. H. Chan, Q. Teng, and J. C. Rajapakse, “Multi- modal graph neural network for early diagnosis of Alzheimer’s disease from sMRI and PET scans,”Computers in Biology and Medicine, vol. 164, p. 107328, 2023

  19. [27]

    Brain effective connectome based on fMRI and DTI data: Bayesian causal learning and assessment,

    A. Bagheri, M. Dehshiri, Y. Bagheri, A. Akhondi-Asl, and B. Nad- jar Araabi, “Brain effective connectome based on fMRI and DTI data: Bayesian causal learning and assessment,”Plos One, vol. 18, no. 8, p. e0289406, 2023

  20. [28]

    BrainNetCNN: Convo- lutional neural networks for brain networks; towards predicting neurodevelopment,

    J. Kawahara, C. J. Brown, S. P . Miller, B. G. Booth, V . Chau, R. E. Grunau, J. G. Zwicker, and G. Hamarneh, “BrainNetCNN: Convo- lutional neural networks for brain networks; towards predicting neurodevelopment,”NeuroImage, vol. 146, pp. 1038–1049, 2017

  21. [29]

    BrainGNN: Interpretable brain graph neural network for fMRI analysis,

    X. Li, Y. Zhou, N. C. Dvornek, M. Zhang, S. Gao, J. Zhuang, D. Scheinost, L. H. Staib, P . Ventola, and J. S. Duncan, “BrainGNN: Interpretable brain graph neural network for fMRI analysis,” in CVPR, 2021, pp. 9260–9270

  22. [30]

    Visualizing data using t-SNE,

    L. Van der Maaten and G. Hinton, “Visualizing data using t-SNE,” Journal of Machine Learning Research, vol. 9, no. 11, 2008

  23. [31]

    Au- tomated anatomical labelling atlas 3,

    E. T. Rolls, C.-C. Huang, C.-P . Lin, J. Feng, and M. Joliot, “Au- tomated anatomical labelling atlas 3,”NeuroImage, vol. 206, p. 116189, 2020

  24. [32]

    J. G. Ratcliffe,Foundations of Hyperbolic Manifolds. Springer, 1994

  25. [33]

    Hyper- bolic geometry,

    J. W. Cannon, W. J. Floyd, R. Kenyon, W. R. Parryet al., “Hyper- bolic geometry,”Flavors of Geometry, vol. 31, pp. 59–115, 1997

  26. [34]

    Poincar ´e embeddings for learning hier- archical representations,

    M. Nickel and D. Kiela, “Poincar ´e embeddings for learning hier- archical representations,” inNeurIPs, 2017, pp. 6338–6347

  27. [35]

    Hyperbolic neural networks,

    O. Ganea, G. B ´ecigneul, and T. Hofmann, “Hyperbolic neural networks,” inNeurIPS, vol. 31, 2018, pp. 5345–5355

  28. [36]

    Poincar ´e kernels for hyperbolic representations,

    P . Fang, M. Harandi, Z. Lan, and L. Petersson, “Poincar ´e kernels for hyperbolic representations,”International Journal of Computer Vision, pp. 1–23, 2023

  29. [37]

    A. A. Ungar,Analytic hyperbolic geometry and Albert Einstein’s special theory of relativity. World Scientific, 2008

  30. [38]

    From pythagoras to einstein: the hyperbolic pythagorean theorem,

    ——, “From pythagoras to einstein: the hyperbolic pythagorean theorem,”Foundations of Physics, vol. 28, no. 8, pp. 1283–1321, 1998

  31. [39]

    Learning dynamic graph rep- resentation of brain connectome with spatio-temporal attention,

    B.-H. Kim, J. C. Ye, and J.-J. Kim, “Learning dynamic graph rep- resentation of brain connectome with spatio-temporal attention,” NeurIPs, vol. 34, pp. 4314–4327, 2021

  32. [40]

    Kernel methods for deep learning,

    Y. Cho and L. Saul, “Kernel methods for deep learning,” in NeurIPS, vol. 22, 2009

  33. [41]

    Random features for large-scale kernel machines,

    A. Rahimi and B. Recht, “Random features for large-scale kernel machines,” inNeurIPS, vol. 20, 2007

  34. [42]

    Generalization properties of learning 14 with random features,

    A. Rudi and L. Rosasco, “Generalization properties of learning 14 with random features,” inNeurIPs, vol. 30, 2017

  35. [43]

    Bach,Learning theory from first principles

    F. Bach,Learning theory from first principles. MIT press, 2024

  36. [44]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” inICLR, 2017

  37. [45]

    Graph attention networks,

    P . Veliˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P . Lio, and Y. Bengio, “Graph attention networks,” inInternational Conference on Learning Representations (ICLR), 2018

  38. [46]

    The Alzheimer’s disease neuroimaging initiative (ADNI): MRI methods,

    C. R. Jack Jr, M. A. Bernstein, N. C. Fox, P . Thompson, G. Alexan- der, D. Harvey, B. Borowski, P . J. Britson, J. L. Whitwell, C. Ward et al., “The Alzheimer’s disease neuroimaging initiative (ADNI): MRI methods,”Journal of Magnetic Resonance Imaging: An Official Journal of ...

  39. [47]

    The autism brain imaging data exchange: Towards a large- scale evaluation of the intrinsic brain architecture in autism,

    A. Di Martino, C.-G. Yan, Q. Li, E. Denio, F. X. Castellanos, K. Alaerts, J. S. Anderson, M. Assaf, S. Y. Bookheimer, M. Dapretto et al., “The autism brain imaging data exchange: Towards a large- scale evaluation of the intrinsic brain architecture in autism,” Molecular Psychi...

  40. [48]

    Reduced default mode network functional connectivity in patients with recurrent major depressive disorder,

    C.-G. Yan, X. Chen, L. Li, F. X. Castellanos, T.-J. Bai, Q.-J. Bo, J. Cao, G.-M. Chen, N.-X. Chen, W. Chenet al., “Reduced default mode network functional connectivity in patients with recurrent major depressive disorder,”Proceedings of the National Academy of Sciences, vol. 1...

  41. [49]

    The ADHD-200 consortium: A model to ad- vance the translational potential of neuroimaging in clinical neu- roscience,

    A.-. Consortium, “The ADHD-200 consortium: A model to ad- vance the translational potential of neuroimaging in clinical neu- roscience,”Frontiers in Systems Neuroscience, vol. 6, p. 62, 2012

  42. [50]

    Support Vector Machine,

    D. A. Pisner and D. M. Schnyer, “Support Vector Machine,” in Machine Learning. Elsevier, 2020, pp. 101–121

  43. [51]

    Random forests,

    L. Breiman, “Random forests,”Machine Learning, vol. 45, pp. 5–32, 2001

  44. [52]

    XGBoost: A scalable tree boosting system,

    T. Chen and C. Guestrin, “XGBoost: A scalable tree boosting system,” inKDD, 2016, pp. 785–794

  45. [53]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, “Attention is all you need,” inNeurIPs, vol. 30, 2017

  46. [54]

    Inductive representation learning on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,”NeurIPS, vol. 30, 2017

  47. [55]

    How powerful are graph neural networks?

    K. Xu, W. Hu, J. Leskovec, and S. Jegelka, “How powerful are graph neural networks?” inICLR, 2019

  48. [56]

    ACTION: Aug- mentation and computation toolbox for brain network analysis with functional MRI,

    Y. Fang, J. Zhang, L. Wang, Q. Wang, and M. Liu, “ACTION: Aug- mentation and computation toolbox for brain network analysis with functional MRI,”NeuroImage, vol. 305, p. 120967, 2025

  49. [57]

    BrainNet Viewer: A network visual- ization tool for human brain connectomics,

    M. Xia, J. Wang, and Y. He, “BrainNet Viewer: A network visual- ization tool for human brain connectomics,”PloS One, vol. 8, no. 7, p. e68910, 2013

  50. [58]

    Subjective cognitive impairment: Functional MRI during a di- vided attention task,

    J. Rodda, T. Dannhauser, D. Cutinha, S. Shergill, and Z. Walker, “Subjective cognitive impairment: Functional MRI during a di- vided attention task,”European Psychiatry, vol. 26, no. 7, pp. 457– 462, 2011

  51. [59]

    Fluid intelligence loss linked to restricted regions of damage within frontal and parietal cortex,

    A. Woolgar, A. Parr, R. Cusack, R. Thompson, I. Nimmo-Smith, T. Torralva, M. Roca, N. Antoun, F. Manes, and J. Duncan, “Fluid intelligence loss linked to restricted regions of damage within frontal and parietal cortex,”Proceedings of the National Academy of Sciences, vol. 107,...

  52. [60]

    Dorsolateral pre- frontal neurons mediate subjective decisions and their variation in humans,

    M. Jamali, B. Grannan, K. Haroush, Z. B. Moses, E. N. Eskandar, T. Herrington, S. Patel, and Z. M. Williams, “Dorsolateral pre- frontal neurons mediate subjective decisions and their variation in humans,”Nature Neuroscience, vol. 22, no. 6, pp. 1010–1020, 2019

  53. [61]

    Resting- state subcortical functional connectivity in HIV-infected patients on long-term cART,

    M. A. Janssen, M. Hinne, R. J. Janssen, M. A. van Gerven, S. C. Steens, B. G ´oraj, P . P . Koopmans, and R. P . Kessels, “Resting- state subcortical functional connectivity in HIV-infected patients on long-term cART,”Brain Imaging and Behavior, vol. 11, pp. 1555– 1560, 2017

  54. [62]

    Altered cerebro-cerebellum resting-state functional connectivity in HIV-infected male patients,

    H. Wang, R. Li, Y. Zhou, Y. Wang, J. Cui, B. A. Nguchu, B. Qiu, X. Wang, and H. Li, “Altered cerebro-cerebellum resting-state functional connectivity in HIV-infected male patients,”Journal of Neurovirology, vol. 24, pp. 587–596, 2018

  55. [63]

    Putamen volume and its clinical and neurological correlates in primary HIV infection,

    P . W. Wright, A. Pyakurel, F. F. Vaida, R. W. Price, E. Lee, J. Pe- terson, D. Fuchs, H. Zetterberg, K. R. Robertson, R. Walteret al., “Putamen volume and its clinical and neurological correlates in primary HIV infection,”Aids, vol. 30, no. 11, pp. 1789–1794, 2016

  56. [64]

    Biochemical changes in the frontal lobe of HIV-infected individuals detected by magnetic resonance spectroscopy,

    D. L ´opez-Villegas, R. E. Lenkinski, and I. Frank, “Biochemical changes in the frontal lobe of HIV-infected individuals detected by magnetic resonance spectroscopy,”Proceedings of the National Academy of Sciences, vol. 94, no. 18, pp. 9854–9859, 1997

  57. [65]

    Empirical optimization of ASL data analysis using an ASL data processing toolbox: ASLtbx,

    Z. Wang, G. K. Aguirre, H. Rao, J. Wang, M. A. Fern ´andez-Seara, A. R. Childress, and J. A. Detre, “Empirical optimization of ASL data analysis using an ASL data processing toolbox: ASLtbx,” Magnetic Resonance Imaging, vol. 26, no. 2, pp. 261–269, 2008

  58. [66]

    Computational radiomics system to decode the radiographic phenotype,

    J. J. Van Griethuysen, A. Fedorov, C. Parmar, A. Hosny, N. Au- coin, V . Narayan, R. G. Beets-Tan, J.-C. Fillion-Robin, S. Pieper, and H. J. Aerts, “Computational radiomics system to decode the radiographic phenotype,”Cancer Research, vol. 77, no. 21, pp. e104– e107, 2017

  59. [67]

    G-mixup: Graph data augmentation for graph classification,

    X. Han, Z. Jiang, N. Liu, and X. Hu, “G-mixup: Graph data augmentation for graph classification,” inInternational Conference on Machine Learning. PMLR, 2022, pp. 8230–8248

Pith tools

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