Pith. sign in

REVIEW 4 major objections 5 minor 64 references

An Explainable Transformer Model for Alzheimer's Disease Detection Using Retinal Imaging

T0 review · 4 major / 5 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 A compact transformer called Retformer, trained from scratch on retinal images, classifies Alzheimer's disease with 92% accuracy on OCT scans and 94% on fundus images, outperforming CNNs and a retinal foundation model by up to 11…

desk verdict A sensible small-data transformer architecture undermined by slice-level leakage and non-blind hyperparameter selection; the headline accuracy claims don't hold as stated. read the letter →

arxiv 2507.04259 v1 pith:QXX76AKK submitted 2025-07-06 cs.LG cs.CV

Saeed Jamshidiha , Alireza Rezaee , Farshid Hajati , Mojtaba Golzan , Raymond Chiong This is my paper · ORCID
classification cs.LGcs.CV
keywords Alzheimer'sdiseasedetectionretinalimagingvisiontransformeropticalcoherencetomographyfundusGrad-CAMexplainableAIsmallmedicaldatasets
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

Retformer is a compact transformer built to detect Alzheimer's disease from two kinds of retinal images—optical coherence tomography (OCT) slices and visible-light fundus photographs. The paper's central claim is that this from-scratch architecture reaches 92% ± 1% accuracy on OCT and 94% ± 8% accuracy on fundus images, beating the strongest comparison models by up to 11 percentage points. The authors attribute the gain to a small-data design: convolutional patch embedding, rotary position embeddings, grouped-query attention, and SwiGLU activations reduce overfitting without pretraining. They also use Grad-CAM to show that the model leans on retinal vasculature, the optic disc, the choroid, and overall retinal thickness, which they read as alignment with known retinal biomarkers of Alzheimer's. If the claims hold, retinal imaging could become a cheaper, non-invasive screening route for a disease whose diagnosis currently depends on costly and less accessible procedures.

What carries the argument

The central object is the Retformer encoder-only transformer. It converts an image into patches through a CNN, adds non-learnable rotary position embeddings (RoPE), passes the sequence through transformer layers that use grouped-query attention (GQA), SwiGLU activations, layer normalization, and two residual connections, and ends in an MLP classifier head. Every choice is aimed at lowering the number of trainable parameters so the model can be trained from scratch on small retinal datasets; the paper's ablation shows the first residual connection, RoPE, and the CNN patch encoder matter most. The explanatory machinery is Grad-CAM, which weights the penultimate feature map by class-specific gradients to produce the localization maps that link model decisions to retinal anatomy.

What would settle it

A concrete test: re-run the OCT comparison with patient-level nested cross-validation, where all slices from a given patient are assigned to the same fold, and report the mean and standard deviation of per-patient accuracy for Retformer and VGG-16; if per-patient accuracy drops to near or below the VGG-16 value, or if slices from the same patient were found to span train and test folds in the current protocol, the paper's central performance claim would collapse.

Watch

Extended reading notes

Core claim

The authors find that a deliberately small, encoder-only vision transformer, with no pretraining, can outperform both ImageNet-pretrained CNNs (EfficientNet-B3, MobileNet, ResNet-50, VGG-16) and the retina-pretrained foundation model RETFound on Alzheimer's classification. On OCT data, Retformer reports accuracy 92% ± 1%, precision 91% ± 2%, sensitivity 90% ± 5%, specificity 91% ± 3%, F1 90% ± 4%, and ROC AUC 96% ± 1%; the best benchmark, VGG-16, reaches 81% ± 4% accuracy. On the fundus dataset it reports 94% ± 8% accuracy and 99% ± 3% AUC. Using Grad-CAM, the paper identifies the optic nerve head and retinal vasculature in fundus images, and the choroid and retinal layers (including RNFL, ganglion, and nuclear layers) in OCT slices, as the decisive regions; a masked-image experiment found that preserving only Grad-CAM-selected regions keeps accuracy at the level of the original images, while random masks degrade it significantly. The authors conclude that the combination of a small, regularized transformer and gradient-based explanation yields both higher accuracy and clinically interpretable decisions.

Load-bearing premise

The load-bearing premise is that the OCT slices are independent samples even though all 1,120 slices come from just 28 patients, with no described patient-level grouping; if the same patient's slices fall into both training and test folds, the reported 92% accuracy and the 11-point lead over baselines are inflated and do not show detection in new patients.

Editorial extensions

If this is right

  • A from-scratch transformer could serve as an AD screening tool using standard OCT and fundus cameras, with accuracy above 90% in the reported settings.
  • Other rare-disease imaging tasks, where labeled data is scarce, could adopt the same toolkit of CNN patch embedding, RoPE, GQA, and SwiGLU to avoid pretraining dependencies.
  • Grad-CAM localization offers a way to connect model predictions to anatomical regions and existing retinal-biomarker literature, potentially guiding future clinical measurement of RNFL, choroid, and vasculature.
  • The claimed parameter efficiency (about 4 million parameters against 42 million for VGG-16) suggests lower computational cost, which matters for community or mobile deployment.

Reading between the lines

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

  • Because the OCT accuracy is computed over 2D slices from only 28 patients, the true patient-level margin over VGG-16 is probably smaller than 11 points; a grouped re-analysis would settle how much of the reported gain is real.
  • The downsampling plateau at 50x50 images suggests the method could transfer to low-cost, low-resolution retinal cameras used in community screening, where clinical-grade OCT is unavailable.
  • The Grad-CAM localization could seed a prospective study: enroll at-risk patients, measure RNFL, choroid, and vasculature in the highlighted regions, and test whether those measurements track cognitive decline.
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 / 5 minor

Summary. The paper proposes Retformer, a compact transformer architecture combining convolutional patch embedding, rotary position embeddings, grouped-query attention, and SwiGLU activations, for detecting Alzheimer's disease from retinal fundus images and OCT image slices. The authors report nested cross-validation experiments on an OCT dataset of 1,120 slices derived from 28 patients and on a combined fundus dataset, claiming accuracy improvements of up to 11% over several baselines including EfficientNet, MobileNet, ResNet-50, VGG-16, and RETFound. They also use Grad-CAM to identify retinal regions important for classification and compare these regions with clinical knowledge of retinal biomarkers of Alzheimer's disease. The central claim is that Retformer outperforms the benchmarks across all reported metrics and provides clinically plausible explanations.

Significance. The proposed architecture is sensibly designed for small medical image datasets, and the paper includes a useful ablation study and a Grad-CAM-based sanity check. If the reported performance were validated with a leakage-free evaluation protocol, the results could be of practical interest for non-invasive AD screening. The paper also makes the used open datasets identifiable, which supports reproducibility. However, the empirical evaluation is currently not trustworthy: the OCT results are based on slice-level splits without patient grouping, which creates a high risk of information leakage and directly undermines the headline comparison against benchmark models.

major comments (4)
  1. [§1.1, §2.2] The OCT evaluation is not patient-independent. Section 1.1 states that the OCT dataset consists of 224 three-dimensional volumes from only 28 patients (14 AD, 14 HC), which are converted into 1,120 two-dimensional slices. Section 2.1 describes nested cross-validation with 3 outer folds but does not mention any grouping by patient. Because adjacent slices from the same 3D volume are highly correlated, random slice-level folds place slices from the same patient in both training and test sets. This allows the model to exploit patient-specific scanning artifacts, anatomy, or imaging conditions rather than generalizable AD biomarkers. Consequently, the reported accuracy of 92% ± 1% on OCT and the 11% accuracy gap over VGG-16 in Table 7 are not reliable evidence for the paper's central claim. A valid evaluation requires patient-level grouping (e.g., GroupKFold), and with only 28 patients the reported standard deviations of 1–5% are implausibly small for patient-level folds.
  2. [§2.1, Tables 3–4] Hyperparameters are selected using a 'simple (non-nested) cross-validation on the entire datasets' before the nested cross-validation evaluation. This is explicitly stated in Section 2.1 and shown in Tables 3 and 4. Selecting hyperparameters on the full dataset leaks information from the eventual test folds into the model configuration, so the nested cross-validation no longer provides an unbiased estimate of generalization performance. This issue is independent of the OCT grouping problem and would need to be corrected by performing hyperparameter selection inside the inner folds of the nested procedure.
  3. [§2.2, Table 8] The p-values in Table 8 are not credible for the stated evaluation design. With only 3 outer folds, any fold-level comparison has at most 2 degrees of freedom for a paired test, and the manuscript's own power calculation in Equations (15)–(16) uses DF = k − 1 = 2. Reporting p < 0.001 for all six metrics requires an implausibly large and consistent t-statistic, and the table does not specify how the test was computed or whether multiple-testing corrections were applied. Moreover, fold-level performance estimates in nested cross-validation are not independent because training sets across folds overlap substantially. The statistical evidence for the claimed superiority of Retformer is therefore not established.
  4. [§2.2 vs. §2.6] There is a direct inconsistency between the reported OCT accuracies. Table 7 reports Retformer accuracy of 92% ± 1% using nested cross-validation, while Table 10 reports a baseline accuracy of 94% in the ablation study using 'a simple 3-fold cross-validation'. The ablation is therefore not comparable to the main evaluation, and the discrepancy suggests that the model's performance depends strongly on the evaluation protocol. This inconsistency weakens the ablation conclusions and, combined with the leakage concerns, indicates that the reported metrics are sensitive to how the experiments are set up.
minor comments (5)
  1. [§2.3, §3] The parameter count of Retformer is inconsistently reported: Section 2.3 states '4 million parameters' while the Discussion states a '7.6M parameter count'. Please clarify which value is correct and reconcile the two numbers.
  2. [Introduction] There is a typo in the Introduction: 'AD-related retinpoathy' should read 'AD-related retinopathy'.
  3. [§2.2, Table 9] The fundus dataset is highly imbalanced (14% AD vs. 86% HC), and only accuracy and AUC are reported. On such a dataset, accuracy can be misleading; the paper should report balanced metrics such as precision, recall, F1, or specificity/sensitivity, as is done for the OCT dataset.
  4. [§2.4, Figure 5] Figure 5 reports accuracy versus image size but does not specify the evaluation protocol (nested CV vs. simple split) or show error bars. Without this information, the downsampling experiment is difficult to interpret.
  5. [§1.2, Equation (3)] The notation in Equation (3) for grouped-query attention is incomplete: the underbrace labels show 'Attention group 1' and 'Attention group g', but intermediate groups are not defined, and the dimensions of the concatenated output are unclear. Please make the notation fully explicit.

Circularity Check

1 steps flagged · score 1.0 of 10

No load-bearing circularity; the accuracy claim is empirical, with only a minor self-referential Grad-CAM sanity check.

  1. other [Section 2.7 (Explainability, Grad-CAM masked-image experiment)]
    "In this setup, we trained a Retformer model on the training data of the outer folds of the nested cross-validation, and used it to generate regions of interest in the test images of the outer folds. ... This suggests that the regions of interest detected by Grad-CAM contribute meaningfully to the model’s performance."

    The Grad-CAM regions of interest are computed from the Retformer model itself, so the masked-image comparison tests the model's saliency maps for consistency with its own decisions rather than independently validating those regions as AD biomarkers. The inference that the highlighted regions are 'relevant for AD diagnosis' is therefore self-referential by construction: the same model generates the masks and is then shown to perform well when trained on those masks. This is a standard explainability sanity check and does not support the paper's headline accuracy comparison, so the overall circularity is minimal.

full rationale

The paper's central claim—that Retformer outperforms EfficientNet, MobileNet, ResNet-50, VGG-16, and RETFound by up to 11% accuracy—is an empirical benchmark result obtained by nested cross-validation; no core equation or fitted parameter reduces to the target result by construction. The architecture choices (RoPE, GQA, SwiGLU, CNN patch embedding) are standard components cited from external work, and the ablations are empirical. Self-citations (References 17-19, involving co-author Chiong) are literature-review items and are not load-bearing. The only self-referential element is the Grad-CAM masked-image sanity check in Section 2.7, which uses the model's own saliency maps to claim the highlighted regions are clinically relevant; this is a secondary interpretability validation, not the source of the reported accuracy. Potential evaluation weaknesses (slice-level OCT leakage from 28 patients and hyperparameters tuned on the full datasets before nested CV) are validity concerns about independence of the benchmark, not circularity of the derivation chain.

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

The paper introduces no new entities. Its central claim depends on the data assumptions above, especially the independence of OCT slices and the integrity of the cross-validation scheme.

free parameters (3)
  • Retformer hyperparameters = D=62/32, patch size 4, layers 1/4, heads 12/2, query groups 3/1 (OCT/fundus)
    Selected via Bayesian optimization against validation performance; data-dependent and could overfit the small datasets.
  • Image size = 50x50 pixels
    Chosen as a tradeoff; downsampling experiments show a plateau, but the 50x50 default is a hand choice.
  • Exclusion of young adult OCT controls = 8 subjects excluded
    Data selection to focus on age-matched comparison; this is a design choice that changes the dataset composition.
assumptions (3)
  • domain assumption 2D OCT slices are independent samples
    Used throughout the nested CV; if false, results are inflated. No patient-level grouping is described.
  • domain assumption Retinal biomarker patterns from clinical studies can be visually compared with Grad-CAM maps
    The XAI discussion assumes a correspondence between model saliency and clinical measurements.
  • standard math Standard transformer components (RoPE, GQA, SwiGLU) behave as reported in prior work
    The architecture relies on published definitions; not derived in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Explainable Transformer Model for Alzheimer's Disease Detection Using Retinal Imaging." pith.science (2026). https://pith.science/paper/QXX76AKK

@misc{pith2026250704259,
  author       = {Pith},
  title        = {Pith review of: An Explainable Transformer Model for Alzheimer's Disease Detection Using Retinal Imaging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QXX76AKK}},
  note         = {Machine review of arXiv:2507.04259}
}
read the original abstract

Alzheimer's disease (AD) is a neurodegenerative disorder that affects millions worldwide. In the absence of effective treatment options, early diagnosis is crucial for initiating management strategies to delay disease onset and slow down its progression. In this study, we propose Retformer, a novel transformer-based architecture for detecting AD using retinal imaging modalities, leveraging the power of transformers and explainable artificial intelligence. The Retformer model is trained on datasets of different modalities of retinal images from patients with AD and age-matched healthy controls, enabling it to learn complex patterns and relationships between image features and disease diagnosis. To provide insights into the decision-making process of our model, we employ the Gradient-weighted Class Activation Mapping algorithm to visualize the feature importance maps, highlighting the regions of the retinal images that contribute most significantly to the classification outcome. These findings are compared to existing clinical studies on detecting AD using retinal biomarkers, allowing us to identify the most important features for AD detection in each imaging modality. The Retformer model outperforms a variety of benchmark algorithms across different performance metrics by margins of up to 11\.

Figures

Figures reproduced from arXiv: 2507.04259 by the authors.

Figure 1
Figure 1. A slice of an OCT image from an AD patient 1.2 Retformer We propose Retformer, a transformer-based model to classify different modalities of retinal images into AD patients and HCs. The Retformer model is an encoder-only transformer that converts input images into a compact latent representation. This representation allows images to be classified using a classifier head, typically a multi-layer perceptron (MLP), whi… view at source ↗
Figure 2
Figure 2. The Retformer model Given the relatively small size of typical medical image datasets in comparison with general machine vision datasets, each component of the Retformer model has been carefully chosen to mitigate the risk of overfitting, ensuring reliable performance even when dealing with small datasets. As a result, we use a non-trainable position encoding scheme, and convolutional layers to embed the input image… view at source ↗
Figure 3
Figure 3. The ROC curve on the OCT data The VGG-16 outperforms the other benchmarks on the OCT dataset. This could be attributed to the simpler structure of the VGG-16 model in comparison to the ResNet-50, EfficientNet-B3, and MobileNet architectures. While the VGG-16 model has a higher parameter count, it is structurally simpler, and is comprised of 16 convolutional layers. The ResNet-50 model, on the other hand, is comprise… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The training accuracies of the Retformer model and the VGG-16 model on the OCT data The training accuracy of the Retformer model is presented in Figure 4a, showing the speed of convergence of the model. As evident in the figure, the Retformer model converges after 65 e…
Figure 5
Figure 5. Figure 5: Accuracy vs. Image Size As evident in the figure, by increasing input image size, model accuracy initially improves due to the increase in the information contained in each image. Afterwards, accuracy plateaus for a while, and afterwards increasing image size leads to …
Figure 6
Figure 6. Figure 6: Training and Validation Accuracy vs. Batch Size 2.6 Ablation Study An ablation study is undertaken in this research to determine the contribution of each building block of the Retformer model to the overall classification, and to compare the impact of the blocks used i…
Figure 7
Figure 7. Figure 7: displays the Grad-CAM output for the Retformer classifier and visualisation of its attention heads, explaining the input image’s classification. Visualising the feature map obtained by the last layer normalisation for the fundus image in Figure 7a using Grad-CAM result…
Figure 8
Figure 8. Figure 8: Visual explanation for Retformer classification of OCT images. Ophthalmologist-labelled data was not available to facilitate a comparison with the Grad-CAM algorithm. However, in order to quantitatively assess the explainability of the Retformer model, we designed the …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

64 extracted references · 51 canonical work pages

  1. [1]

    Association, A. et al. 2016 alzheimer’s disease facts and figures. Alzheimer’s & Dementia 12, 459–509 (2016)

    work page 2016

  2. [2]

    A., Jack Jr, C

    Sperling, R. A., Jack Jr, C. R. & Aisen, P. S. Testing the right target and right drug at the right stage. Sci. translational medicine 3 (2011)

    work page 2011

  3. [3]

    Dubois, B. et al. Preclinical alzheimer’s disease: definition, natural history, and diagnostic criteria.Alzheimer’s & Dementia 12, 292–323 (2016)

    work page 2016

  4. [4]

    Sperling, R. A. et al. Toward defining the preclinical stages of alzheimer’s disease: Recommendations from the national institute on aging-alzheimer’s association workgroups on diagnostic guidelines for alzheimer’s disease. Alzheimer’s & dementia 7, 280–292 (2011)

    work page 2011

  5. [5]

    K., Altinas, Ö., Tokay, T

    Iseri, P. K., Altinas, Ö., Tokay, T. & Yüksel, N. Relationship between cognitive impairment and retinal morphological and visual functional abnormalities in alzheimer disease. J. neuro-ophthalmology 26, 18–24 (2006)

    work page 2006

  6. [6]

    & Durmus, M

    Kirbas, S., Turkyilmaz, K., Anlar, O., Tufekci, A. & Durmus, M. Retinal nerve fiber layer thickness in patients with alzheimer disease. J. Neuro-Ophthalmology 33, 58–61 (2013)

    work page 2013

  7. [7]

    Kromer, R. et al. Detection of retinal nerve fiber layer defects in alzheimer’s disease using sd-oct. Front. psychiatry 5, 22 (2014)

    work page 2014

  8. [8]

    D., Visser, P

    den Haan, J., Verbraak, F. D., Visser, P. J. & Bouwman, F. H. Retinal thickness in alzheimer’s disease: a systematic review and meta-analysis. Alzheimer’s & Dementia: Diagn. Assess. & Dis. Monit. 6, 162–170 (2017). 17/20

    work page 2017

Show all 64 references
  1. [9]

    Neurocognitive assessment and retinal thickness alterations in alzheimer disease: is there a correlation? J

    Cipollini, V .et al. Neurocognitive assessment and retinal thickness alterations in alzheimer disease: is there a correlation? J. Neuro-Ophthalmology 40, 370–377 (2020)

    2020

  2. [10]

    Y ., Choi, E

    Shin, J. Y ., Choi, E. Y ., Kim, M., Lee, H. K. & Byeon, S. H. Changes in retinal microvasculature and retinal layer thickness in association with apolipoprotein e genotype in alzheimer’s disease. Sci. Reports 11, 1847 (2021)

    2021

  3. [11]

    Y .et al

    Cheung, C. Y .et al. Retinal imaging in alzheimer’s disease. J. Neurol. Neurosurg. & Psychiatry92, 983–994 (2021)

    2021

  4. [12]

    Y .-l., Ikram, M

    Cheung, C. Y .-l., Ikram, M. K., Chen, C. & Wong, T. Y . Imaging retina to study dementia and stroke.Prog. retinal eye research 57, 89–107 (2017)

    2017

  5. [13]

    Retinal biomarkers in alzheimer’s disease and mild cognitive impairment: A systematic review and meta-analysis

    Ge, Y .-J.et al. Retinal biomarkers in alzheimer’s disease and mild cognitive impairment: A systematic review and meta-analysis. Ageing Res. Rev. 69, 101361 (2021)

    2021

  6. [14]

    A foundation model for generalizable disease detection from retinal images

    Zhou, Y .et al. A foundation model for generalizable disease detection from retinal images. Nature 622, 156–163 (2023)

    2023

  7. [15]

    & Fatima, A

    Badar, M., Haris, M. & Fatima, A. Application of deep learning for retinal image analysis: A review. Comput. Sci. Rev. 35, 100203 (2020)

    2020

  8. [16]

    & Javadzadeh, A

    Ashayeri, H., Jafarizadeh, A., Yousefi, M., Farhadi, F. & Javadzadeh, A. Retinal imaging and alzheimer’s disease: a future powered by artificial intelligence. Graefe’s Arch. for Clin. Exp. Ophthalmol. 1–13 (2024)

    2024

  9. [17]

    A., Luo, S

    Ebrahimighahnavieh, M. A., Luo, S. & Chiong, R. Deep learning to detect alzheimer’s disease from neuroimaging: A systematic literature review. Comput. methods programs biomedicine 187, 105242 (2020)

    2020

  10. [18]

    Ebrahimi, A., Luo, S., Chiong, R., Initiative, A. D. N. et al. Deep sequence modelling for alzheimer’s disease detection using mri. Comput. Biol. Medicine 134, 104537 (2021)

    2021

  11. [19]

    & Chiong, R

    Ebrahimi-Ghahnavieh, A., Luo, S. & Chiong, R. Transfer learning for alzheimer’s disease detection on mri images. In 2019 IEEE International Conference on Industry 4.0, Artificial Intelligence, and Communications Technology (IAICT), 133–138 (IEEE, 2019)

    2019

  12. [20]

    Tian, J. et al. Modular machine learning for alzheimer’s disease classification from retinal vasculature. Sci. Reports 11, 238 (2021)

    2021

  13. [21]

    Tan, M. & Le, Q. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, 6105–6114 (PMLR, 2019)

    2019

  14. [22]

    Y .et al

    Cheung, C. Y .et al. A deep learning model for detection of alzheimer’s disease based on retinal photographs: a retrospective, multicentre case-control study. The Lancet Digit. Heal. 4, e806–e815 (2022)

    2022

  15. [23]

    & Sun, J

    He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, 770–778 (2016)

    2016

  16. [24]

    Wisely, C. E. et al. Convolutional neural network to identify symptomatic alzheimer’s disease using multimodal retinal imaging. Br. J. Ophthalmol. 106, 388–395 (2022)

    2022

  17. [25]

    Wang, X. et al. Machine learning based on optical coherence tomography images as a diagnostic tool for alzheimer’s disease. CNS Neurosci. & Ther. 28, 2206–2217 (2022)

    2022

  18. [26]

    & Lesage, F

    Corbin, D. & Lesage, F. Assessment of the predictive potential of cognitive scores from retinal images and retinal fundus metadata via deep learning using the clsa database. Sci. Reports 12, 5767 (2022)

    2022

  19. [27]

    & Zisserman, A

    Simonyan, K. & Zisserman, A. Very deep convolutional networks for large-scale image recognition. CoRR abs/1409.1556 (2014)

    2014 arXiv

  20. [28]

    Yousefzadeh, N. et al. Neuron-level explainable ai for alzheimer’s disease assessment from fundus images. Sci. Reports 14, 7710 (2024)

    2024

  21. [29]

    & El-Sharkawy, M

    Sinha, D. & El-Sharkawy, M. Thin mobilenet: An enhanced mobilenet architecture. In 2019 IEEE 10th annual ubiquitous computing, electronics & mobile communication conference (UEMCON), 0280–0285 (IEEE, 2019)

    2019

  22. [30]

    K., Lim, Y

    Kim, H. K., Lim, Y . J., Sunwoo, M. H.et al. Efficient deep retinal fundus image-based network for alzheimer’s disease diagnosis using mobile device applications. IEEE Access (2024)

    2024

  23. [31]

    & Ben Ayed, I

    Silva-Rodríguez, J., Chakor, H., Kobbi, R., Dolz, J. & Ben Ayed, I. A foundation language-image model of the retina (flair): encoding expert knowledge in text supervision. Med. Image Analysis 99, 103357, DOI: https://doi.org/10.1016/j. media.2024.103357 (2025)

    2025

  24. [32]

    Du, J. et al. Ret-clip: A retinal image foundation model pre-trained with clinical diagnostic reports. In Linguraru, M. G. et al. (eds.) Medical Image Computing and Computer Assisted Intervention – MICCAI 2024, 709–719 (Springer Nature Switzerland, Cham, 2024). 18/20

    2024

  25. [33]

    & Arora, S

    Vij, R. & Arora, S. A systematic survey of advances in retinal imaging modalities for alzheimer’s disease diagnosis. Metab. Brain Dis. 37, 2213–2243 (2022)

    2022

  26. [34]

    Shamshad, F. et al. Transformers in medical imaging: A survey. Med. image analysis 88, 102802 (2023)

    2023

  27. [35]

    Hussain, T. et al. Effresnet-vit: A fusion-based convolutional and vision transformer model for explainable medical image classification. IEEE Access (2025)

    2025

  28. [36]

    Alam, T. et al. An integrated approach using yolov8 and resnet, seresnet & vision transformer (vit) algorithms based on roi fracture prediction in x-ray images of the elbow. Curr. Med. Imaging 20, e15734056309890 (2024)

    2024

  29. [37]

    & Liu, F

    Dai, Y ., Gao, Y . & Liu, F. Transmed: Transformers advance multi-modal medical image classification.Diagnostics 11, 1384 (2021)

    2021

  30. [38]

    Touvron, H. et al. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, 10347–10357 (PMLR, 2021)

    2021

  31. [39]

    Monajatipoor, M. et al. Berthop: An effective vision-and-language model for chest x-ray disease diagnosis. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, 725–734 (Springer, 2022)

    2022

  32. [40]

    Valanarasu, J. M. J., Oza, P., Hacihaliloglu, I. & Patel, V . M. Medical transformer: Gated axial-attention for medical image segmentation. In Medical image computing and computer assisted intervention–MICCAI 2021: 24th international conference, Strasbourg, France, September 2...

    2021

  33. [41]

    A., Marhoon, H

    Hussain, T., Shouno, H., Mohammed, M. A., Marhoon, H. A. & Alam, T. Dcssga-unet: Biomedical image segmentation with densenet channel spatial and semantic guidance attention. Knowledge-Based Syst. 314, 113233 (2025)

    2025

  34. [42]

    N., Ibrahim, M

    Asim, M. N., Ibrahim, M. A., Imran Malik, M., Dengel, A. & Ahmed, S. Circ-locnet: A computational framework for circular rna sub-cellular localization prediction. Int. J. Mol. Sci. 23, 8221 (2022)

    2022

  35. [43]

    Khan, S. U. R., Asim, M. N., V ollmer, S. & Dengel, A. Robust & precise knowledge distillation-based novel context-aware predictor for disease detection in brain and gastrointestinal. arXiv preprint arXiv:2505.06381 (2025)

    2025

  36. [44]

    Dwivedi, R. et al. Explainable ai (xai): Core ideas, techniques, and solutions. ACM Comput. Surv. 55, 1–33 (2023)

    2023

  37. [45]

    & Khan, S

    Hekmat, A., Zuping, Z., Bilal, O. & Khan, S. U. R. Differential evolution-driven optimized ensemble network for brain tumor detection. Int. J. Mach. Learn. Cybern. 1–26 (2025)

    2025

  38. [46]

    Selvaraju, R. R. et al. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, 618–626 (2017)

    2017

  39. [47]

    Vaswani, A. et al. Attention is all you need. Adv. neural information processing systems 30 (2017)

    2017

  40. [48]

    Dosovitskiy, A. et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (2021)

    2021

  41. [49]

    Su, J. et al. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing 568, 127063 (2024)

    2024

  42. [50]

    Ainslie, J. et al. GQA: Training generalized multi-query transformer models from multi-head checkpoints. In Bouamor, H., Pino, J. & Bali, K. (eds.) Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 4895–4901, DOI: 10.18653/v1/2023.emnlp-ma...

    2023 doi

  43. [51]

    Glu variants improve transformer

    Shazeer, N. Glu variants improve transformer. arXiv preprint arXiv:2002.05202 (2020)

    2020 arXiv

  44. [52]

    & Bernard, J

    Bissig, D., Zhou, C., Le, V . & Bernard, J. A practical approach to functional optical coherence tomography shows abnormal retinal responses in alzheimer’s disease [dataset]. dryad. (2020). https://doi.org/10.5061/dryad.msbcc2ftc

    2020 doi

  45. [53]

    Popovic, N. et al. Topological characterization of the retinal microvascular network visualized by portable fundus camera- effects of chronic disease (trend2) database (1.1) [dataset]. zenodo. (2023)

    2023

  46. [54]

    Cen, L.-P. et al. Automatic detection of 39 fundus diseases and conditions in retinal photographs using deep neural networks. Nat. communications 12, 4828 (2021)

    2021

  47. [55]

    L., Kiros, J

    Ba, J. L., Kiros, J. R. & Hinton, G. E. Layer normalization (2016). 1607.06450

    2016 arXiv

  48. [56]

    D., Wijewardena, P

    Keles, F. D., Wijewardena, P. M. & Hegde, C. On the computational complexity of self-attention. In International Conference on Algorithmic Learning Theory, 597–619 (PMLR, 2023)

    2023

  49. [57]

    & Jégou, H

    Touvron, H., Cord, M., El-Nouby, A., Verbeek, J. & Jégou, H. Three things everyone should know about vision transformers. In European Conference on Computer Vision, 497–515 (Springer, 2022). 19/20

    2022

  50. [58]

    Ke, X., Cao, W. & Lv, F. Relationship between complexity and precision of convolutional neural networks. In2017 2nd International Symposium on Advances in Electrical, Electronics and Computer Engineering (ISAEECE 2017), 325–329 (Atlantis Press, 2017)

    2017

  51. [59]

    & Aghaeipoor, F

    Kashefi, R., Barekatain, L., Sabokrou, M. & Aghaeipoor, F. Explainability of vision transformers: A comprehensive review and new perspectives. arXiv preprint arXiv:2311.06786 (2023)

    2023 arXiv

  52. [60]

    Deng, J. et al. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, 248–255 (Ieee, 2009)

    2009

  53. [61]

    & Zeng, W

    Wang, G., Zhao, Y ., Tang, C., Luo, C. & Zeng, W. When shift operation meets vision transformer: An extremely simple alternative to attention mechanism. In Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, 2423–2430 (2022)

    2022

  54. [62]

    Lee, C. S. & Apte, R. S. Retinal biomarkers of alzheimer disease. Am. journal ophthalmology 218, 337–341 (2020)

    2020

  55. [63]

    Budu-Aggrey, A. et al. The relationship between open angle glaucoma, optic disc morphology and alzheimer’s disease: a mendelian randomization study. medRxiv 2020–08 (2020)

    2020

  56. [64]

    Bambo, M. P. et al. Analysis of optic disk color changes in alzheimer’s disease: A potential new biomarker. Clin. Neurol. Neurosurg. 132, 68–73 (2015). 20/20

    2015

Pith tools

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