Pith. sign in

REVIEW 4 major objections 6 minor 41 references

Not Only Grey Matter: OmniBrain for Robust Multimodal Classification of Alzheimer's Disease

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

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

Pith's one-line read OmniBrain is a multimodal Alzheimer's classifier that fuses grey-matter MRI, radiomics, gene expression, and clinical metadata through cross-attention with modality masking, reporting 92.15% accuracy on ANMerge and 70.37% zero-shot…

desk verdict Useful multimodal AD benchmark with a genuinely external test, but the headline zero-shot number rests on an unverified preprocessing assumption that the authors need to check. read the letter →

arxiv 2507.20872 v1 pith:T7ZQDC2V submitted 2025-07-28 cs.CV cs.AIcs.LG

Ahmed Sharshar , Yasser Ashraf , Tameem Bakr , Salma Hassan , Hosam Elgendy , Mohammad Yaqub , Mohsen Guizani This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords Alzheimer'sdiseasemultimodalclassificationcross-attentionfusionmodalitydropoutradiomicsgreymatterMRIgeneexpressionzero-shotgeneralization
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

OmniBrain is a multimodal classifier for Alzheimer's disease that combines four sources of evidence: grey-matter MRI maps, radiomics from 32 brain regions, blood gene-expression profiles, and clinical metadata such as MMSE and APOE. The paper argues that no single modality is enough, and that a cross-attention fusion trained with modality masking can handle real-world missing data while still outperforming unimodal and prior multimodal systems. On the ANMerge cohort the full model reaches 92.15% accuracy, and trained on ANMerge alone it transfers to the MRI-only ADNI cohort with 70.37% accuracy. The authors also show that the model's attention concentrates on the hippocampus, parahippocampal gyrus, and thalamus, and that its top tabular features include MMSE, APOE, and genes previously linked to oxidative stress and mitochondrial dysfunction. The paper's central claim is that a single framework can achieve accuracy, cross-dataset generalization, missing-modality tolerance, and explainability at once.

What carries the argument

The argument is carried by a cross-attention fusion block that combines a visual stream and a tabular stream. The visual stream encodes grey-matter maps with a pre-trained anatomical contrastive backbone (AnatCL or y-Aware InfoNCE), while the tabular stream processes radiomics, gene expression, and clinical metadata with an FT-Transformer. A Modality-Aware Attention Masking scheme randomly drops entire modalities during training, so the attention layer learns to suppress empty channels at inference instead of imputing them. Radiomics are extracted from 32 brain regions segmented by SynthSeg, and the gene panel is narrowed from 230 literature-curated genes to 139 via an ANOVA F-test. Explainability comes from Grad-CAM on the grey-matter maps and SHAP on the tabular features, which together tie the model's decisions to known neuropathological targets.

What would settle it

Compare SynthSeg-derived volumes against manual or FreeSurfer segmentations on a held-out set of ADNI scans that includes severely atrophied AD cases; if the volumetric concordance is poor or the SynthSeg QC metric fails disproportionately on AD, then the 70.37% zero-shot number cannot be attributed to the fusion model. A second check: train the exact same pipeline on ANMerge but preprocess ADNI through the same four-stage pipeline (registration, skull-stripping, N4 correction, normalization) and see whether the accuracy changes; a large shift would show the result depends on the untested 'comparable corrections' assumption.

Watch

Extended reading notes

Core claim

The central discovery is that integrating voxel-level grey matter with radiomics, gene expression, and clinical metadata in a single cross-attention model under modality dropout yields substantially higher classification accuracy than any single modality or prior multimodal approach, especially on the hard AD vs. MCI and MCI vs. CTL boundaries. With all four signals, OmniBrain reports 92.15% ± 2.4 accuracy on ANMerge, setting new state-of-the-art results on those two tasks by up to 4.8 percentage points. Trained only on ANMerge, it achieves 70.37% ± 2.7 accuracy on 300 ADNI scans, exceeding the best unimodal baseline by 17.2 percentage points. The architecture does not impute missing inputs; instead, modality-aware attention masking during training teaches the model to re-weight the available modalities, so inference degrades gracefully when genetic and clinical data are absent. Explainability analyses with Grad-CAM and SHAP show the model attends to the hippocampus, parahippocampal gyrus, and thalamus, and weights MMSE, APOE, and established AD-related genes most heavily.

Load-bearing premise

The load-bearing premise is that SynthSeg segments the 32 brain regions accurately on both ANMerge and ADNI, yet it is validated only against ANMerge volume references; if segmentation degrades on ADNI acquisitions or on atrophied brains, the radiomics features and the cross-dataset comparison both become unreliable.

Editorial extensions

If this is right

  • Combining imaging with genetic and clinical signals raises accuracy by about 15 percentage points over MRI-plus-radiomics alone on ANMerge, so imaging-only Alzheimer's classifiers are leaving useful diagnostic signal on the table.
  • Because the model is trained with modality masking, it can be deployed in settings where genetic or clinical data are unavailable, with only a 2–3 point drop in in-domain accuracy.
  • The zero-shot result on ADNI indicates that a model trained on one cohort can be applied to another scanner protocol and population without retraining, at least for three-way CTL/MCI/AD classification.
  • If the explainability findings hold, the model's saliency maps and feature rankings could serve as a screening tool that points clinicians to the same regions and biomarkers they would examine manually.

Reading between the lines

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

  • The 17.2-point margin over the best unimodal baseline on ADNI is partly a comparison against a weak radiomics-only baseline (53.15%); a stronger imaging baseline would likely shrink that margin, so the headline generalization number should be read as 'better than the tested baselines,' not as an absolute ceiling.
  • A natural test of the masking strategy is to compare it head-to-head with generative imputation methods on the same missing-modality protocol; the paper does not include such a comparison, and it is plausible that imputation helps when the missing channel is informative.
  • Because SynthSeg is validated only against ANMerge volumetric references, an immediate extension is to quantify segmentation quality on ADNI and check whether per-case segmentation errors correlate with the model's misclassifications; if they do, radiomics features are the bottleneck rather than the fusion mechanism.
  • The claim that LSM3 and CCDC72 are potential novel AD biomarkers is speculative until replicated on independent cohorts; the paper itself flags only that they have not been previously linked.
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 OmniBrain, a multimodal framework that combines T1-weighted MRI grey-matter maps (encoded by the AnatCL or y-Aware InfoNCE foundation models), PyRadiomics features from SynthSeg-based parcellation of 32 brain regions, blood-based gene-expression features, and clinical metadata (MMSE, APOE) via an FT-Transformer tabular encoder and a cross-attention fusion module with modality masking. The model is trained on the ANMerge dataset using patient-level five-fold cross-validation and evaluated in-domain and via zero-shot transfer to the ADNI dataset. The authors report 92.15% three-class accuracy on ANMerge and 70.37% on ADNI (MRI-only) without retraining, claim state-of-the-art results on the AD vs. MCI and MCI vs. CTL subtasks, and provide Grad-CAM and SHAP explainability analyses.

Significance. If the results hold, OmniBrain would be a valuable contribution to multimodal Alzheimer's classification, addressing the practically important problem of inference with missing modalities and providing external validation on an independent cohort. The methodological choices are largely appropriate: patient-level group splits prevent subject leakage, gene selection is nested inside cross-validation folds, and the ADNI experiment is a genuinely held-out test set. The in-domain ANMerge result is supported by the reported protocol. However, the ADNI zero-shot claim currently rests on an unverified preprocessing equivalence and on a table that marks genes and metadata as present even though ADNI lacks them, so the exact configuration that yields 70.37% is unclear. The explainability analyses are plausible but not quantitatively validated against clinical annotations. With the identified issues resolved, the paper would provide a useful benchmark for robust multimodal diagnosis.

major comments (4)
  1. [§3.1 and §6] The statement that 'No preprocessing steps were required for the ADNI dataset, since it already includes comparable corrections' is asserted without quantitative support. Every imaging feature that OmniBrain consumes—SynthSeg segmentations and derived PyRadiomics features, CAT12 grey-matter maps, and AnatCL embeddings—is sensitive to the exact preprocessing chain (registration, skull-stripping, bias correction, intensity normalization). ANMerge is processed through a four-stage pipeline, but ADNI is not. If the two datasets enter the model in different feature spaces, the reported 70.37% accuracy and the +17.2% margin over the radiomics-only baseline conflate population/disease-related domain shift with an unmeasured preprocessing covariate shift. The authors should either run ADNI through the identical preprocessing pipeline or provide quantitative evidence (e.g., feature-distribution comparisons, segmentation QC metrics on ADNI) that the features are comparable.
  2. [§6, Table 3] In Table 3, the ADNI rows that yield the headline results (68.78% and 70.37%) mark Genes and Meta as present (●), yet the text repeatedly states that ADNI lacks gene expression and clinical metadata and that the missing-modality strategy masks these inputs at inference. This is internally inconsistent and makes it impossible to tell which modality set actually produced 70.37%. If the reported numbers are from a model with Genes and Meta fully masked, the table must mark those modalities as absent (❍) for all ADNI rows. If, instead, some form of imputation or partial retention was used, that must be described explicitly. As written, the claim of 'MRI-only' zero-shot transfer is not supported by the table.
  3. [§3, ADNI Dataset] The construction of the ADNI test set is under-specified. The paper says only that 'We sampled 300 different MRI scans of patients (100 per class)' from the ADNI cohort. This leaves open whether each subject contributes exactly one scan, whether any subject appears multiple times, how the 100-per-class sample was drawn from a presumably larger and imbalanced cohort, and whether any acquisition-phase or site restrictions applied. Because the reported accuracy is on a class-balanced subset, it is not directly comparable to numbers reported on naturalistic cohorts, and the result may not be reproducible without the exact sampling protocol. The authors should report the sampling strategy, inclusion/exclusion criteria, and ideally also provide results on an unselected or designated ADNI test partition.
  4. [§6, Table 4] Table 4 compares OmniBrain against prior methods that use different datasets (ADNI vs. ANMerge), different modality sets, and different binary tasks without protocol matching. While the '+4.8%' claim over the prior ANMerge-based methods (Maddalena 2023, Hassan 2024) is internally consistent, the table's presentation implies global state-of-the-art across the listed ADNI studies, which are not directly comparable because they use different cohorts and class definitions. The authors should either restrict the SOTA claim to the ANMerge dataset with matched evaluation protocols or add a clear caveat that the comparisons are not protocol-matched.
minor comments (6)
  1. [§3.3] The text contains a typo: 'ANOV A F-test' should be 'ANOVA F-test'.
  2. [§6, paragraph after Table 5] The paragraph states '85.29% accuracy (85.39% F1) for AD vs CTL,' but Table 5 reports 86.29% accuracy; please correct the discrepancy.
  3. [§6, Table 2 and Table 3] The legend 'Light-green rows = missing-modality runs' does not indicate which modality is dropped in each row. Please specify the masked modality or modality combination for each missing-modality run.
  4. [§4] The sentence 'T1-weighted MRI scans are processed into GM maps to extract radiomics features' is imprecise, since radiomics features are extracted from the SynthSeg parcellation of the full brain (including white matter and subcortical structures), not from the grey-matter maps alone. Please rephrase.
  5. [Figure 5] The caption for Figure 5 describes axial MRI slices showing GM and WM segmentations, but the figure appears to be a bar chart of SHAP feature importances. Please update the caption to match the figure content.
  6. [§3.2] The text lists '137 features' and then '37 matrix-based texture features' as if separate; it would be clearer to state the full breakdown of first-order, shape, and texture features that sum to the total radiomics count.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: OmniBrain's reported predictions are evaluated on external and held-out data, with feature selection nested inside folds.

full rationale

The paper's central claims are empirical benchmark results, not derived quantities. The in-domain ANMerge numbers come from a patient-level 5-fold cross-validation, and the generalization claim is a zero-shot evaluation on ADNI: the model is trained only on ANMerge and then tested on 300 ADNI scans without retraining, so the 70.37% ADNI accuracy is not fitted or defined in terms of the ADNI labels. Gene selection via ANOVA F-test is explicitly performed inside folds to avoid label leakage. SynthSeg and the radiomics pipeline are validated against ANMerge volumetric references, and AnatCL/y-Aware are externally pretrained backbones used as feature extractors, not as circular inputs to the classification objective. The one same-author citation, Hassan et al. 2024 [15], appears as a comparison baseline in Table 4 and in related work; it is an externally published result and is not load-bearing for OmniBrain's architecture or for the claim that the proposed fusion works. The paper's assumption that ADNI 'already includes comparable corrections' and the apparent inconsistency in Table 3 (Genes and Meta marked present in ADNI rows even though ADNI lacks them) are correctness and generalization-risk concerns, but they do not make any prediction equal to a fitted input or reduce a derivation to its own definition. No circular step can be exhibited from the text; the honest finding is no significant circularity.

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

The central claims rest on dataset label validity, SynthSeg segmentation quality across datasets, transferability of pretrained MRI encoders, and the effectiveness of modality-dropout training. These are reasonable domain assumptions for an applied ML paper, but none is verified beyond the reported accuracy numbers.

free parameters (4)
  • Gene selection threshold and score weights = p<0.01, weights 70% F-score and 30% negative log p, top 139 of 230 genes
    Hand-selected in Section 3.3; changing the threshold or the top-k count changes the tabular feature set and all reported accuracies.
  • Number of cross-attention heads = 4
    Section 5 states the model is fine-tuned with four cross-attention heads to limit overfitting; this is a hand-chosen design decision.
  • Training hyperparameters = learning rate 1e-5, weight decay 5e-4, batch size 64, up to 50 epochs, patience 5
    Given in Section 5.1; these choices affect the reported means and standard deviations.
  • Focal loss class weights = Inversely proportional to class frequency
    Section 5.1 uses weighted focal loss to address class imbalance; the exact weighting scheme affects the accuracy and recall numbers.
assumptions (4)
  • domain assumption ANMerge and ADNI diagnostic labels for CTL, MCI, and AD are correct and comparable across the two cohorts.
    Section 3 treats both datasets as ground truth without independent adjudication; label noise or different diagnostic criteria would weaken the generalization claim.
  • domain assumption SynthSeg segmentation is accurate on both ANMerge and ADNI, despite being validated only against ANMerge reference volumes.
    Section 3.1 derives all radiomics features from SynthSeg regions; incorrect masks on ADNI would corrupt the external test features.
  • domain assumption AnatCL and y-Aware InfoNCE pretrained on large adult T1 MRI corpora transfer to pathological, older Alzheimer's brains.
    Section 4 uses these foundation models as feature extractors for grey matter; no fine-tuning on pathological anatomy is described beyond the ANMerge task.
  • domain assumption Modality dropout during training is sufficient for the attention module to handle entirely absent modalities at inference.
    Section 4 asserts this behavior; the only supporting evidence is the single ADNI transfer experiment, with no theoretical or ablative proof that attention weights do not depend on the presence of all training modalities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Not Only Grey Matter: OmniBrain for Robust Multimodal Classification of Alzheimer's Disease." pith.science (2026). https://pith.science/paper/T7ZQDC2V

@misc{pith2026250720872,
  author       = {Pith},
  title        = {Pith review of: Not Only Grey Matter: OmniBrain for Robust Multimodal Classification of Alzheimer's Disease},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T7ZQDC2V}},
  note         = {Machine review of arXiv:2507.20872}
}
abstract

Alzheimer's disease affects over 55 million people worldwide and is projected to more than double by 2050, necessitating rapid, accurate, and scalable diagnostics. However, existing approaches are limited because they cannot achieve clinically acceptable accuracy, generalization across datasets, robustness to missing modalities, and explainability all at the same time. This inability to satisfy all these requirements simultaneously undermines their reliability in clinical settings. We propose OmniBrain, a multimodal framework that integrates brain MRI, radiomics, gene expression, and clinical data using a unified model with cross-attention and modality dropout. OmniBrain achieves $92.2 \pm 2.4\%$accuracy on the ANMerge dataset and generalizes to the MRI-only ADNI dataset with $70.4 \pm 2.7\%$ accuracy, outperforming unimodal and prior multimodal approaches. Explainability analyses highlight neuropathologically relevant brain regions and genes, enhancing clinical trust. OmniBrain offers a robust, interpretable, and practical solution for real-world Alzheimer's diagnosis.

Figures

Figures reproduced from arXiv: 2507.20872 by the authors.

Figure 1
Figure 1. Multimodal diagnostic pipeline for dementia classification. MRI scans are processed to extract grey matter maps and radiomics [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Overview of the multimodal pre-processing pipeline for [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Pre-processing pipeline for MRI data for standardiza [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Grad-CAM attention overlays in axial, coronal and sagittal views for (a) AD, (b) CTL and (c) MCI, showing the network’s [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Axial slices of T1-weighted MRI showing processed im [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 39 canonical work pages

  1. [1]

    Adarsh, G

    V . Adarsh, G. R. Gangadharan, Ugo Fiore, and Pierluigi Zanetti. Multimodal classification of alzheimer’s disease and mild cognitive impairment using custom mkscddl kernel over cnn with transparent decision-making for explainable diagnosis. Scientific Reports, 14(1):1774, 2024. 2

    work page 2024

  2. [2]

    Classification of alzheimer disease on imaging modal- ities with deep cnns using cross-modal transfer learning

    Karim Aderghal, Alexander Khvostikov, Andrei Krylov, Jenny Benois-Pineau, Karim Afdel, and Gwenaelle Cathe- line. Classification of alzheimer disease on imaging modal- ities with deep cnns using cross-modal transfer learning. In 2018 IEEE 31st International Symposium on Computer- Based Medical Systems (CBMS) , pages 345–350, 2018. 2, 7

    work page 2018

  3. [3]

    M. A. Aghdam, S. Bozdag, F. Saeed, et al. Machine-learning models for alzheimer’s disease diagnosis using neuroimag- ing data: survey, reproducibility, and generalizability evalu- ation. Brain Informatics, 12(8), 2025. 1

    work page 2025

  4. [4]

    Alatrany, Wajid Khan, Amir Hussain, Hoshang Kolivand, and Dhiya Al-Jumeily

    Ameer S. Alatrany, Wajid Khan, Amir Hussain, Hoshang Kolivand, and Dhiya Al-Jumeily. An explainable machine learning approach for alzheimer’s disease classification. Sci- entific Reports, 14(1):2637, 2024. 2

    work page 2024

  5. [5]

    Alzheimer’s Disease Facts and Figures

    Alzheimer’s Association. Alzheimer’s Disease Facts and Figures. Alzheimer’s Association, 2024. 1

    work page 2024

  6. [6]

    Multiuser Wirelessly Powered Backscatter Communications: Nonlinearity, Waveform Design and SINR-Energy Tradeoff

    Tobias Backstr ¨om, Roger Johansson, Tomas L ¨ofstedt, and Charlotte Forsell. An efficient 3d deep convolutional net- work for alzheimer’s disease diagnosis using mr images. arXiv preprint arXiv:1811.01356, 2018. 2

    work page Pith review arXiv 2018

  7. [7]

    Anatomical foundation models for brain mris, 2024

    Carlo Alberto Barbano, Matteo Brunello, Benoit Dufumier, and Marco Grangetto. Anatomical foundation models for brain mris, 2024. 5

    work page 2024

  8. [8]

    Greve, Oula Puonti, Axel Thielscher, Koen Van Leemput, Bruce Fischl, Adrian V

    Benoit Billot, Douglas N. Greve, Oula Puonti, Axel Thielscher, Koen Van Leemput, Bruce Fischl, Adrian V . Dalca, Juan Eugenio Iglesias, and ADNI. Synthseg: Seg- mentation of brain mri scans of any contrast and resolu- tion without retraining. Medical Image Analysis, 86:102789,

    work page

Show all 41 references
  1. [9]

    Anmerge: A comprehensive and accessible alzheimer’s disease patient- level dataset

    Clemens Birkenbihl, Sarah Westwood, Lei Shi, Ana Nevado- Holgado, Eric Westman, Simon Lovestone, AddNeuroMed Consortium, and Gerhard Hofmann-Apitius. Anmerge: A comprehensive and accessible alzheimer’s disease patient- level dataset. Journal of Alzheimer’s Disease , 79(1):423–...

    2021

  2. [10]

    Automated de- tection of Alzheimer’s disease: a multi-modal approach with 3D MRI and amyloid PET

    Giovanna Castellano, Arianna Esposito, Eufemia Lella, Gioacchino Montanaro, and Gennaro Vessio. Automated de- tection of Alzheimer’s disease: a multi-modal approach with 3D MRI and amyloid PET. Scientific Reports, 14(1):5210,

  3. [11]

    Contrastive learning with continuous proxy meta-data for 3d mri classification, 2021

    Benoit Dufumier, Pietro Gori, Julie Victor, Antoine Grigis, Michel Wessa, Paolo Brambilla, Pauline Favre, Mircea Polosan, Colm McDonald, Camille Marie Piguet, and Edouard Duchesnay. Contrastive learning with continuous proxy meta-data for 3d mri classification, 2021. 5

    2021

  4. [12]

    Ronald A. Fisher. Statistical Methods for Research Workers. Oliver and Boyd, Edinburgh, 1st edition, 1925. 4

    1925

  5. [13]

    Thompson, Florian Kurth, and Eileen Luders

    Christian Gaser, Robert Dahnke, Paul M. Thompson, Florian Kurth, and Eileen Luders. Cat: a computational anatomy toolbox for the analysis of structural mri data. GigaScience, 13:giae049, 2024. 4

    2024

  6. [14]

    Multimodal attention-based deep learning for Alzheimer’s disease diagnosis

    Maxim Golovanevsky, Carsten Eickhoff, and Ritambhara Singh. Multimodal attention-based deep learning for Alzheimer’s disease diagnosis. Journal of the American Medical Informatics Association , 29(12):2014–2022, 2022. 2

    2014

  7. [15]

    Mindsets: Multi-omics integration with neuroimaging for dementia subtyping and effective temporal study, 2024

    Salma Hassan, Dawlat Akaila, Maryam Arjemandi, Vijay Papineni, and Mohammad Yaqub. Mindsets: Multi-omics integration with neuroimaging for dementia subtyping and effective temporal study, 2024. 2, 7

    2024

  8. [16]

    Henr ´ıquez and Nicol ´as Araya

    Pedro A. Henr ´ıquez and Nicol ´as Araya. Multimodal Alzheimer’s disease classification through ensemble deep random vector functional link neural network. PeerJ Com- puter Science, 10, 2024. 2

    2024

  9. [17]

    Itcfn: In- complete triple-modal co-attention fusion network for mild cognitive impairment conversion prediction

    Xiangyang Hu, Xiangyu Shen, Yifei Sun, Xuhao Shan, Wen- wen Min, Liyilei Su, Xiaomao Fan, Ahmed Elazab, Rui- quan Ge, Changmiao Wang, and Xiaopeng Fan. Itcfn: In- complete triple-modal co-attention fusion network for mild cognitive impairment conversion prediction. arXiv prepr...

    2025 arXiv

  10. [18]

    Huang, A

    X. Huang, A. J. Jannu, Z. Song, N. Jury-Garfe, C. A. Lasagna-Reeves, Alzheimer’s Disease Neuroimaging Initia- tive, T. S. Johnson, K. Huang, and J. Zhang. Predicting alzheimer’s disease subtypes and understanding their molec- ular characteristics in living patients with transc...

  11. [19]

    Explaining graph convolutional network predictions for clin- icians – an xai approach to ad classification

    Md Islam, Samir Ghanvatkar, Jorge Ruiz, and Rahul Rajan. Explaining graph convolutional network predictions for clin- icians – an xai approach to ad classification. Frontiers in Artificial Intelligence, 6:1334613, 2023. 2

    2023

  12. [20]

    Classification of alzheimer’s disease based on multiple cluster dense convolutional networks

    Fan Li, Mingxia Liu, and Alzheimer’s Disease Neuroimag- ing Initiative. Classification of alzheimer’s disease based on multiple cluster dense convolutional networks. Computer- ized Medical Imaging and Graphics , 70:101–110, 2018. 2, 7

    2018

  13. [21]

    Cascaded multi-modal mixing transformers for alzheimer’s disease classification with incomplete data

    Linfeng Liu, Siyu Liu, Lu Zhang, Xuan Vinh To, Fatima Nasrallah, and Shekhar S Chandra. Cascaded multi-modal mixing transformers for alzheimer’s disease classification with incomplete data. arXiv preprint arXiv:2210.00255 ,

    arXiv

  14. [22]

    To, and Andy C

    Lei Liu, Shuwei Liu, Lei Zhang, Xavier V . To, and Andy C. Cheung. Cascaded multi-modal mixing transformers for alzheimer’s disease classification with incomplete data.Neu- roImage, 277:120267, 2023. 2

    2023

  15. [23]

    A unified approach to interpreting model predictions

    Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Advances in Neural In- formation Processing Systems 30, pages 4765–4774. Curran Associates, Inc., 2017. 5

    2017

  16. [24]

    A blood gene expression marker of early alzheimer’s disease

    Katie Lunnon, Martina Sattlecker, Simon J Furney, Giovanni Coppola, Andrew Simmons, Petroula Proitsi, Michelle K Lupton, Anbarasu Lourdusamy, Caroline Johnston, Hilkka Soininen, et al. A blood gene expression marker of early alzheimer’s disease. Journal of Alzheimer’s Disease,...

    2013

  17. [25]

    Classifying alzheimer’s disease using mris and transcriptomic data

    Lucia Maddalena, Irene Granata, Mariarosaria Giordano, Mauro Manzo, and Mario Guarracino. Classifying alzheimer’s disease using mris and transcriptomic data. In Proceedings of the 15th International Joint Confer- ence on Biomedical Engineering Systems and Technologies (BIOIMAG...

    2022

  18. [26]

    Integrating omics and imaging data for alzheimer’s disease classification

    Lucia Maddalena, Ilaria Granata, Maurizio Giordano, Mario Manzo, and Mario Rosario Guarracino. Integrating omics and imaging data for alzheimer’s disease classification. Arti- ficial Intelligence in Medicine, 123:102186, 2022. 2, 7

    2022

  19. [27]

    Integrating different data modalities for the classification of alzheimer’s disease stages

    Lucia Maddalena, Irene Granata, Mariarosaria Giordano, Mauro Manzo, and Maria Rosaria Guarracino. Integrating different data modalities for the classification of alzheimer’s disease stages. SN Computer Science, 4(249):1–14, 2023. 2, 7

    2023

  20. [28]

    Alzheimer’s disease blood biomarkers associated with neuroinflammation as therapeutic targets for early personalized intervention

    Sher Li Oh, Meikun Zhou, Eunice WM Chin, Gautami Amarnath, Chee Hoe Cheah, Eyleen LK Ng, Kok Pin, and Keng-Hwee Chiam. Alzheimer’s disease blood biomarkers associated with neuroinflammation as therapeutic targets for early personalized intervention. Frontiers in Digital Health...

    2022

  21. [29]

    Dementia

    World Health Organization. Dementia. 2023. 1

    2023

  22. [30]

    Multi-modality im- age classification with deep convolutional neural network for alzheimer’s disease prediction

    Zhiyong Pan, Xin Liang, Hongjun Tang, Zhenyu Li, Sheng- wei Zhong, Lei Sun, and Lijun Liu. Multi-modality im- age classification with deep convolutional neural network for alzheimer’s disease prediction. In 2018 IEEE International Conference on Bioinformatics and Biomedicine (...

    2018

  23. [31]

    Petersen, Paul S

    Ronald C. Petersen, Paul S. Aisen, Laurel A. Beckett, Michael C. Donohue, Anthony C. Gamst, Dominic J. Har- vey, Clifford R. Jr Jack, William J. Jagust, Leslie M. Shaw, Arthur W. Toga, John Q. Trojanowski, and Michael W. Weiner. Alzheimer’s disease neuroimaging initiative (adn...

    2010

  24. [32]

    Comparing automated subcortical volume estimation methods; amygdala volumes estimated by fsl and freesurfer have poor consistency

    Patrick Sadil and Martin A Lindquist. Comparing automated subcortical volume estimation methods; amygdala volumes estimated by fsl and freesurfer have poor consistency. Tech- nical report, Wiley Online Library, 2024. 4

    2024

  25. [33]

    Schellenberg, R

    G. Schellenberg, R. Corces, A. DeStefano, L. Farrer, A. Goate, B. Kunkle, F. Leung, E. Marcora, R. Mayeux, and B. Vardarajan. List of ad loci and genes with genetic evi- dence compiled by adsp gene verification committee, 2023. Accessed: 2025-04-06. 4, 7

    2023

  26. [34]

    Grad-cam: Visual explanations from deep networks via gradient-based localization

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE in- ternational conference on computer vision , pages 618–626,

  27. [35]

    Multimodal stacked deep poly- nomial network for alzheimer’s disease diagnosis

    Jun Shi, Xinyue Zheng, Yan Li, Qingjie Zhang, Shihui Ying, Gang Pan, and Tianzi Jiang. Multimodal stacked deep poly- nomial network for alzheimer’s disease diagnosis. Pattern Recognition, 88:271–284, 2019. 2, 7

    2019

  28. [36]

    Explainability of 3d cnns for fmri-based ad classification using grad-cam

    Byung Song, Satoshi Yoshida, and for the Alzheimer’s Dis- ease Neuroimaging Initiative. Explainability of 3d cnns for fmri-based ad classification using grad-cam. PLoS ONE, 19 (5):e0303278, 2024. 2

    2024

  29. [37]

    Joost J. M. van Griethuysen, Andriy Fedorov, Chintan Parmar, Ahmed Hosny, Nicole Aucoin, Vivek Narayan, Regina G. H. Beets-Tan, Jean-Christophe Fillion-Robin, Steve Pieper, and Hugo J. W. L. Aerts. Computational ra- diomics system to decode the radiographic phenotype. Can- cer...

    2017

  30. [38]

    Deep multimodal learning with missing modality: A survey, 2024

    Renjie Wu, Hu Wang, Hsiang-Ting Chen, and Gustavo Carneiro. Deep multimodal learning with missing modality: A survey, 2024. 5

    2024

  31. [39]

    Flex-moe: Modeling arbitrary modality com- bination via the flexible mixture-of-experts

    Sukwon Yun, Inyoung Choi, Jie Peng, Yangfan Wu, Jingx- uan Bao, Qiyiwen Zhang, Jiayi Xin, Qi Long, and Tian- long Chen. Flex-moe: Modeling arbitrary modality com- bination via the flexible mixture-of-experts. arXiv preprint arXiv:2410.08245, 2024. 2

    arXiv

  32. [40]

    Azin Zarei, Aryan Keshavarz, Ehsan Jafari, Reza Nemati, Arash Farhadi, Ali Gholamrezanezhad, Habib Rostami, and Majid Assadi. Automated classification of Alzheimer’s dis- ease, mild cognitive impairment, and cognitively normal pa- tients using 3D convolutional neural network a...

    2024

  33. [41]

    Identification of alzheimer’s disease and mild cog- nitive impairment using networks constructed based on mul- tiple morphological brain features

    Weikai Zheng, Zhisong Yao, Yanan Xie, Jiawei Fan, and Bing Hu. Identification of alzheimer’s disease and mild cog- nitive impairment using networks constructed based on mul- tiple morphological brain features. Biological Psychiatry: Cognitive Neuroscience and Neuroimaging, 3(1...

Pith tools

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