Pith. sign in

REVIEW 5 major objections 5 minor 66 references

PiPViT: Patch-based Visual Interpretable Prototypes for Retinal Image Analysis

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Vision transformer classifier for retinal OCT learns prototypes that localize biomarkers such as fluid and drusen, approximating lesion extent using only image-level labels and without pixel supervision.

desk verdict Reasonable incremental prototype method with a promising direction, but its lesion-extent claim rests on a nonstandard AP metric that needs to be redone before the paper can be trusted. read the letter →

arxiv 2506.10669 v2 pith:2KGQPSNP submitted 2025-06-12 cs.CV cs.AI

Marzieh Oghbaie , Teresa Araújo , Hrvoje Bogunović This is my paper
classification cs.CVcs.AI
keywords interpretabilityprototype-basedmodelsvisiontransformersopticalcoherencetomographyretinaldiseaseclassificationweaklysupervisedlocalizationcontrastivelearning
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

PiPViT is a prototype-based classifier for retinal OCT images that aims to produce, from image-level labels alone, activation maps in pixel space that correspond to clinically meaningful biomarkers such as intraretinal fluid, drusen, and retinal atrophy. The paper argues that by treating the feature maps of a vision transformer as prototypes, applying contrastive and sparsity objectives, and pre-training at multiple input resolutions, the model can approximate both the presence and the extent of lesions without any pixel-level supervision. A sympathetic reader would care because clinicians need to know not just that a model says 'disease' but where in the scan the disease is and how much tissue it affects, and prototype methods have previously produced fragments or mismatched visualizations. The paper reports competitive classification accuracy across four datasets and a quantitative drusen-localization experiment, arguing that the learned prototypes are semantically aligned with real biomarkers.

What carries the argument

The central mechanism is the use of the ViT's spatial feature maps as prototypes. From an input image, a ViT encoder produces a $W'\times H'\times D$ feature tensor; each of the $D$ channels is a prototype, and the max-pooled value of a channel, after a softmax over channels, gives a presence score $p\in[0,1]^D$. A fully connected layer with non-negative, sparsity-regularized weights turns these scores into a transparent class scoring sheet. Contrastive alignment, tanh diversity, and KoLeo regularization shape the feature space during a multi-resolution pre-training step, and the same losses are carried into fine-tuning along with the classification loss.

What would settle it

Take a held-out set of OCT B-scans with expert pixel-level segmentations of intraretinal fluid and geographic atrophy. If PiPViT's prototypes, thresholded at any level, produce detections whose IoU with the expert segmentations is not significantly above random overlap, the claim that activation maps approximate lesion extent fails. A simpler check: re-run the paper's drusen experiment with a standard object-detection AP protocol (fixed IoU thresholds) instead of the scale-factor box protocol; an AP near zero would refute the localization claim.

Watch

Extended reading notes

Core claim

The central discovery, as the authors state it, is that a ViT-based prototypical network can learn robust prototypes that are human-interpretable and approximate lesion extent while using only image-level labels. Each feature map of the ViT output is treated as a prototype; the max-pooled value measures the presence of that prototype, and a sparse non-negative linear layer maps these presence scores to class predictions. Multi-resolution contrastive pre-training, together with alignment, tanh, and KoLeo losses, shapes a feature space in which a prototype's spatial activation corresponds to the location and spread of a biomarker in the input. The paper's quantitative evaluation on drusen bounding boxes, and its qualitative examples for fluid, geographic atrophy, and epiretinal membrane, are offered as evidence that the activation maps delineate lesion extent rather than mere discriminative fragments.

Load-bearing premise

The model is never shown pixel-level annotations, so the claim that a prototype's activation map reflects the true location and extent of a biomarker rests on the assumption that max-pooled channel activations, resized to the input, are faithful spatial indicators; the paper tests this only for drusen.

Editorial extensions

If this is right

  • Clinicians could see, for a given diagnosis, which prototype fired and where the corresponding biomarker sits in the B-scan, without needing a separate segmentation model.
  • The approach transfers the interpretability burden from post-hoc heatmaps to the model itself, so the explanation is the decision path.
  • Multi-resolution pre-training should make the model more robust to different OCT devices and resolutions, a practical requirement for deployment across clinics.
  • If the prototype-to-extent mapping holds across biomarkers, the same architecture could be applied to other imaging modalities that need lesion localization from weak labels.

Reading between the lines

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

  • Editorial extension: A straightforward next test would be to evaluate the same prototype activation maps against pixel-level segmentations for fluid and geographic atrophy in addition to drusen; if those hold up, the localization claim is much stronger than the single-biomarker test in the paper suggests.
  • Editorial extension: The sparsity-regularized scoring sheet invites a second use: after training, the linear weights can be inspected to see which prototypes are shared across diseases, potentially revealing shared biomarkers such as fluid appearing in both RVO and DME.
  • Editorial extension: The nonstandard drusen AP protocol (boxes scaled by a factor rather than IoU-matched detections) makes the absolute AP numbers hard to compare with object-detection literature; re-running with standard detection metrics would either strengthen or temper the reported localization advantage.
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

5 major / 5 minor

Summary. The paper proposes PiPViT, a prototype-based vision transformer for retinal OCT B-scan classification and weakly supervised lesion localization. The model treats each ViT feature map as a prototype, uses max pooling to obtain presence scores, and applies a sparse linear layer for classification. It is pre-trained with a multi-resolution contrastive objective (alignment, tanh, and KoLeo losses) and fine-tuned with classification plus self-supervised losses. Experiments cover four OCT classification datasets and a drusen localization experiment on OCT5K, where prototype activation maps are compared against expert bounding-box annotations. The paper reports competitive classification accuracy relative to several prototype-based and black-box baselines, and it claims that the learned prototypes approximate lesion extent despite using only image-level labels.

Significance. If the core claim is supported, PiPViT would be a useful step toward inherently interpretable models for retinal imaging, because the prototypes align with clinically meaningful biomarkers and convey spatial extent rather than only coarse attention. The authors provide code, report bootstrap confidence intervals for classification, and use an external drusen annotation set (OCT5K), which are strengths. However, the quantitative evidence for the lesion-extent claim is fragile: the AP protocol in §3.3 is nonstandard, the paper does not establish that the OCT5K evaluation images were disjoint from the classifier's training set, and the reported AP values do not show PiPViT outperforming ProtoPNet. The classification results themselves are sound and show the model is competitive, but the interpretability claim, which is the paper's main novelty, is not yet convincingly demonstrated.

major comments (5)
  1. [§3.3, Table 2] The average-precision protocol used for drusen detection is not the standard object-detection AP. Standard AP ranks detections by confidence at a fixed IoU threshold and computes the area under the precision-recall curve as the confidence threshold varies. Here, instead, a box is placed around each highlighted region and a scale factor s in [0.2,10] is swept; precision and recall are computed at each scale. This changes the detector geometry rather than the detection threshold, and there is no confidence ranking or fixed IoU criterion. Consequently, the reported values (e.g., 0.227 for PiPViT) are not comparable to AP scores in the detection literature, and they do not measure localization accuracy in the usual sense. The authors should either re-run the evaluation with a standard protocol (e.g., thresholding activation maps at multiple confidence levels, ranking detections by score, and computing mAP at IoU 0.5) or explicitly rename the metric (e.g., scale-averaged precision) and avoid claims of detection AP.
  2. [§3.3, data overlap] OCT5K is a subset of the Kermany dataset, and the binary drusen/Normal classifier is trained on a subset of Kermany. The paper does not state that the 105 OCT5K B-scans used for evaluation were excluded from the training subset of that classifier. If any of those images were seen during training, the reported localization numbers are optimistically biased. This is load-bearing because the entire lesion-extent claim rests on this experiment. The authors must explicitly describe the sample-level split and confirm that no OCT5K image was used for training the binary classifier or for any pre-training/fine-tuning step.
  3. [Table 2, §4.3] The paper's own quantitative results do not support the claim that PiPViT's prototypes localize drusen better than other prototype methods. ProtoPNet achieves AP-top1 0.272 and AP-top2 0.273, both higher than PiPViT's 0.227 and 0.25. The text acknowledges this but dismisses it by arguing that ProtoPNet's maps are coarse-grained; that is a qualitative argument. Under the reported metric, ProtoPNet is superior. The paper needs to either explain why the metric favors ProtoPNet while the method is still preferable, or provide additional quantitative evidence (e.g., standard AP, IoU coverage, or per-image localization error) that supports the superiority claim.
  4. [§4.3.1, Table 3] The sensitivity analysis in Table 3 is reported without confidence intervals or significance tests, and the differences are modest (e.g., 0.176 vs 0.195 vs 0.227). The claim that increasing patch size and input resolution 'enhanced localization' is based on single runs of the nonstandard AP metric. Additionally, the analysis is performed only for drusen; the paper claims the method approximates lesion extent for multiple biomarkers (IRF, GA, RVO), but no quantitative localization evidence is provided for those. The authors should temper the claim or add experiments for at least one other biomarker, and they should report variability (e.g., multiple runs or bootstrap over B-scans) for the AP values.
  5. [§3.4, classification loss] There is an inconsistency in the experimental setup: Section 3.4 states 'For classification, we used focal loss,' while the paragraph on pre-training says 'In both models, we used negative log-likelihood loss.' It is unclear which loss is used for the final classifier in PiPViT and PiP-Net. The authors should clarify the exact classification objective, since focal loss versus NLL can affect both accuracy and the learned prototypes.
minor comments (5)
  1. [§4.2] There is a typo: 'PiViT' appears in a bullet point and should be 'PiPViT'.
  2. [§3.4] The text says 'pertaining' where 'pre-training' is intended (two occurrences).
  3. [§3.3] The description of how 'highlighted regions' are defined is incomplete. The paper says a box is placed around each highlighted region in the activation map, but it does not specify how regions are extracted (e.g., threshold on activation values, connected components, blob detection). This is needed for reproducibility.
  4. [§2.3] The sparsity regularizer notation 'log((p w_c)^n + 1)' is ambiguous: it should clarify whether the logarithm is applied element-wise or to the product, and how the sparsity order n affects the optimization.
  5. [§3.1] For the OCT5K dataset, the paper states 'we utilized 105 B-scans containing 1132 annotated drusen boxes,' but it does not state how these 105 B-scans were selected (e.g., random, all available with annotations, stratified across patients). A short description would aid reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PiPViT's prototypes are trained with image-level labels only, and the lesion-extent evaluation uses external OCT5K bounding-box annotations, so the central claim is not an artifact of fitting or self-citation.

full rationale

The paper's derivation chain does not reduce to its inputs. Prototype activations are produced by a ViT feature space optimized with contrastive losses (Eqs. 1-4) and an image-level classification loss (Eq. 5); no pixel-level supervision is used. The quantitative semantic-alignment result (Table 2) is computed by placing boxes around activation-map maxima and comparing them to OCT5K drusen bounding boxes, which are external annotations not used in training. The AP protocol in Section 3.3 sweeps a box scale rather than a confidence threshold, which weakens the metric's standard interpretability, but it is applied identically to all methods and does not make the reported PiPViT numbers forced by construction. The paper's self-citations ([13], [44]) supply a private dataset and baseline/ViT context, not a load-bearing uniqueness argument; the core method is an adaptation of external work (PiP-Net [43]) with a ViT backbone and multi-resolution pre-training. One residual risk is not circularity: OCT5K is a subset of Kermany, and the paper does not explicitly state that the 105 evaluation B-scans were excluded from the binary classifier's training subset; if they were not, the localization numbers would be optimistic, but that would be data leakage, not a definitional fit. The acknowledged limitation that the scoring mechanism can overemphasize non-clinical regions further shows the claim is conditional rather than tautological.

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

The method contributes a network architecture and training recipe; it does not derive constants from theory. The main hand-chosen elements are the loss weighting, the sparsity order, learning rates, and the input resolution/patch size. The domain assumptions are that image-level labels suffice to learn spatially meaningful feature maps and that activation maxima can be read as pixel-space lesion extent. The OCT5K annotations are taken as ground truth.

free parameters (4)
  • Loss weights lambda_A, lambda_T, lambda_C, lambda_KoLeo = Not reported (selected as 'reasonable values')
    Section 3.4 states these were chosen without extensive search, but exact values are not given, so all quantitative results depend on unreported constants.
  • Sparsity order n = Not reported (default is 2 from [43])
    Section 2.3 suggests higher values but does not state the value used in experiments, affecting the sparsity of the scoring sheet and potentially prototype selection.
  • Learning rate(s) after grid search = Not reported
    Section 3.4: 'tuned learning rates via grid search', but the selected values are never listed, which is a key training detail.
  • Input resolution and patch size (multi-resolution schedule) = 384x384 input, patch 32; pre-trained at 224, 384, 512
    Section 3.4 and Table 3: the final setting is 384/32; this is a hand-chosen architectural choice that the sensitivity analysis shows affects AP substantially.
assumptions (4)
  • domain assumption Max-pooled activations of feature maps, projected to input space, represent the spatial extent of the corresponding biomarker.
    Used in Section 2.1 to define prototype presence scores and in Section 4.3 to generate boxes around highlighted activation regions. If activations are not spatially coherent in input space, the lesion-extent claim fails.
  • domain assumption Image-level labels provide enough information to learn spatially meaningful prototypes.
    The method uses only image-level supervision (Section 2 and 3.1); the claimed alignment with biomarkers depends on this rather than on pixel-level labels.
  • domain assumption The OCT5K drusen bounding boxes are a reliable ground truth for localization quality.
    Table 2 treats these boxes as true positives/negatives; no annotation-quality assessment is provided.
  • standard math The adopted contrastive losses (alignment, tanh, KoLeo) behave as claimed and are inherited from prior work.
    Equations 1-3 are taken from [43] and [45] and are not re-derived; the paper relies on their known properties.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PiPViT: Patch-based Visual Interpretable Prototypes for Retinal Image Analysis." pith.science (2026). https://pith.science/paper/2KGQPSNP

@misc{pith2026250610669,
  author       = {Pith},
  title        = {Pith review of: PiPViT: Patch-based Visual Interpretable Prototypes for Retinal Image Analysis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2KGQPSNP}},
  note         = {Machine review of arXiv:2506.10669}
}
read the original abstract

Background and Objective: Prototype-based methods improve interpretability by learning fine-grained part-prototypes; however, their visualization in the input pixel space is not always consistent with human-understandable biomarkers. In addition, well-known prototype-based approaches typically learn extremely granular prototypes that are less interpretable in medical imaging, where both the presence and extent of biomarkers and lesions are critical. Methods: To address these challenges, we propose PiPViT (Patch-based Visual Interpretable Prototypes), an inherently interpretable prototypical model for image recognition. Leveraging a vision transformer (ViT), PiPViT captures long-range dependencies among patches to learn robust, human-interpretable prototypes that approximate lesion extent only using image-level labels. Additionally, PiPViT benefits from contrastive learning and multi-resolution input processing, which enables effective localization of biomarkers across scales. Results: We evaluated PiPViT on retinal OCT image classification across four datasets, where it achieved competitive quantitative performance compared to state-of-the-art methods while delivering more meaningful explanations. Moreover, quantitative evaluation on a hold-out test set confirms that the learned prototypes are semantically and clinically relevant. We believe PiPViT can transparently explain its decisions and assist clinicians in understanding diagnostic outcomes. Github page: https://github.com/marziehoghbaie/PiPViT

Figures

Figures reproduced from arXiv: 2506.10669 by the authors.

Figure 3
Figure 3. Given an input image 𝑥 ∈ ℝ𝑊 ×𝐻 , where 𝑊 and 𝐻 represent width and height, PiPViT utilizes a ViT backbone to extract low-level feature maps 𝑧 ∈ ℝ𝑊 ′×𝐻′×𝐷. Here, 𝑊 ′ and 𝐻′ (denoted as 𝑆 in [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 1
Figure 1. An RVO sample with IRF as the most prominent biomarker. (a) Post-hoc explanations for both black-box and non-prototypical approaches highlight lesions to some extent, yet they fail to fully explain the model’s decision. (b) The top-5 activated prototypes of ProtoPNet are repetitive, mostly highlighting the same spatial region in the input image, which also does not correspond to IRF [PITH_FULL_IMAGE:figures/full_fi… view at source ↗
Figure 2
Figure 2. Activation map of the prototype extracted by PiP￾Net corresponding to fluid in DME cases. The examples illustrate varying fluid pocket sizes, where the prototype fails to approximate the extent of the affected areas. Oghbaie et al.: Preprint submitted to Elsevier Page 11 of 10 [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figures from the paper (8 more)
Figure 3
Figure 3. Figure 3: The pipeline of our proposed PiPViT: The ViT encoder extracts patch representations from the input image at a given resolution. Pre-training is conducted using three different image resolutions, with consistent patch size and adaptively resized positional embeddings. T…
Figure 4
Figure 4. Figure 4: Drusen detection is framed as an object detection task by mapping the drusen prototype into pixel space to highlight potential drusen regions (purple box). A true positive occurs when the ground truth overlaps the highlighted region (green); a false negative when a dru…
Figure 6
Figure 6. Figure 6: Visual explanation of different models’ decisions for a DME sample in the OPTIMA5C. (a) Heatmaps generated by post-hoc approaches applied to black-box and non-prototypical models. (b) Top-5 highly activated prototypes of PiPViT vs. baseline prototypical methods (ProtoP…
Figure 5
Figure 5. Figure 5: Classification results on the test set, with 95% confidence intervals from 1000 bootstrap replicates shown as error bars. To facilitate comparison between methods, different patterns are used for each subgroup: black-box models are represented with “xx”, non-prototypic…
Figure 9
Figure 9. Figure 9: Visualization of the top-5 highly activated prototypes learned by PiPViT for class GA in OPTIMA5C. PiPViT effec￾tively captures key GA biomarkers, including hyperreflective areas (column within the red rectangle) and retinal layer atrophy (column within the yellow rect…
Figure 7
Figure 7. Figure 7: Visual explanation of different models’ decisions for an ERM sample in OCTDL. (a) Heatmaps generated by post-hoc approaches applied to black-box and non-prototypical models. (b) Top-5 highly activated prototypes of PiPViT vs. baseline prototypical methods: PiPViT ident…
Figure 11
Figure 11. Figure 11: Visual explanation of different models’ decisions for a drusen sample from OCT5K. (a) Heatmaps generated by post-hoc approaches applied to black-box and non-prototypical models. (b) Highly activated prototype of PiPViT vs. baseline prototypical methods. Oghbaie et al.…
Figure 12
Figure 12. Figure 12: Visual comparison of the top-1 highly activated drusen prototype across different pre-training settings of PiPViT on OCT5K. We study the impact of image resolution, patch size, and incorporation of multi-resolution pre-training on the semantic quality of the learned d…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

66 extracted references · 53 canonical work pages

  1. [1]

    Quantifying attention flow in transformers

    Samira Abnar and Willem Zuidema. Quantifying attention flow in transformers. arXivpreprint arXiv:2005.00928, 2020

    arXiv 2005

  2. [2]

    bioRxiv,pages2023– 03, 2023

    Mustafa Arikan, James Willoughby, Sevim Ongun, Ferenc Sallo, AndreaMontesel,HendAhmed,AhmedHagag,MariusBook,Henrik Faatz,MariaVittoriaCicinelli,etal.Oct5k:Adatasetofmulti-disease andmulti-gradedannotationsforretinallayers. bioRxiv,pages2023– 03, 2023

    work page 2023

  3. [3]

    Interpretable detection of epiretinal membrane from optical coherence tomography with deep neural networks

    Murat Seçkin Ayhan, Jonas Neubauer, Mehmet Murat Uzel, Faik Gelisken, and Philipp Berens. Interpretable detection of epiretinal membrane from optical coherence tomography with deep neural networks. Scientific Reports, 14(1):8484, 2024

    work page 2024

  4. [4]

    Deeplearning predicts hip fracture using confounding patient and healthcare vari- ables

    MarcusABadgeley,JohnRZech,LukeOakden-Rayner,BenjaminS Glicksberg, Manway Liu, William Gale, Michael V McConnell, BethanyPercha,ThomasMSnyder,andJoelTDudley. Deeplearning predicts hip fracture using confounding patient and healthcare vari- ables. NPJdigital medicine, 2(1):31, 2019

    work page 2019

  5. [5]

    A case-based interpretable deep learning model for classification of mass lesions in digital mammography.Nature Machine Intelligence, 3(12):1061– 1070, 2021

    Alina Jade Barnett, Fides Regina Schwartz, Chaofan Tao, Chaofan Chen, Yinhao Ren, Joseph Y Lo, and Cynthia Rudin. A case-based interpretable deep learning model for classification of mass lesions in digital mammography.Nature Machine Intelligence, 3(12):1061– 1070, 2021

    work page 2021

  6. [6]

    Flexivit: One model for all patch sizes

    LucasBeyer,PavelIzmailov,AlexanderKolesnikov,MathildeCaron, SimonKornblith,XiaohuaZhai,MatthiasMinderer,MichaelTschan- nen, Ibrahim Alabdulmohsin, and Filip Pavetic. Flexivit: One model for all patch sizes. InProceedings of the IEEE/CVF Conference on ComputerVisionandPatternRecognition,pages14496–14506,2023

    work page 2023

  7. [7]

    B- cos alignment for inherently interpretable cnns and vision transform- ers

    Moritz Böhle, Navdeeppal Singh, Mario Fritz, and Bernt Schiele. B- cos alignment for inherently interpretable cnns and vision transform- ers. IEEETransactionsonPatternAnalysisandMachineIntelligence, 2024

    work page 2024

  8. [8]

    Transformer interpretability beyond attention visualization

    Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 782– 791, 2021

    work page 2021

Show all 66 references
  1. [9]

    This looks like that: deep learning for in- terpretable image recognition

    Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for in- terpretable image recognition. Advances in neural information processingsystems, 32, 2019

    2019

  2. [10]

    Randaugment:Practicalautomateddataaugmentationwithareduced search space

    Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment:Practicalautomateddataaugmentationwithareduced search space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702–703, 2020

    2020

  3. [11]

    Aiforradiographic covid-19 detection selects shortcuts over signal

    AlexJDeGrave,JosephDJanizek,andSu-InLee. Aiforradiographic covid-19 detection selects shortcuts over signal. Nature Machine Intelligence, 3(7):610–619, 2021

    2021

  4. [12]

    This actually looks like that: Proto-bagnets for local and global interpretability-by-design

    Kerol Djoumessi, Bubacarr Bah, Laura Kühlewein, Philipp Berens, and Lisa Koch. This actually looks like that: Proto-bagnets for local and global interpretability-by-design. arXiv preprint arXiv:2406.15168, 2024. Oghbaie et al.:Preprint submitted to ElsevierPage 8 of 10 PiPViT

    2024 arXiv

  5. [13]

    Pretrained deep 2.5 d models for efficient predictive modeling from retinal oct: A pinnacle study report

    Taha Emre, Marzieh Oghbaie, Arunava Chakravarty, Antoine Rivail, Sophie Riedl, Julia Mai, Hendrik PN Scholl, Sobha Sivaprasad, Daniel Rueckert, Andrew Lotery, et al. Pretrained deep 2.5 d models for efficient predictive modeling from retinal oct: A pinnacle study report. In In...

    2023

  6. [14]

    Attention to lesion: Lesion-aware convolu- tionalneuralnetworkforretinalopticalcoherencetomographyimage classification

    LeyuanFang,ChongWang,ShutaoLi,HosseinRabbani,Xiangdong Chen, and Zhimin Liu. Attention to lesion: Lesion-aware convolu- tionalneuralnetworkforretinalopticalcoherencetomographyimage classification. IEEE transactions on medical imaging, 38(8):1959– 1970, 2019

    1959

  7. [15]

    Fan Gan, Fei-Peng Wu, and Yu-Lin Zhong. Artificial intelligence method based on multi-feature fusion for automatic macular edema (me)classificationonspectral-domainopticalcoherencetomography (sd-oct) images.Frontiersinneuroscience, 17:1097291, 2023

    2023

  8. [16]

    Early and intermediate age-related macular degeneration: update and clinical review

    Alfredo García-Layana, Francisco Cabrera-López, José García- Arumí, Lluís Arias-Barquet, and José M Ruiz-Moreno. Early and intermediate age-related macular degeneration: update and clinical review. Clinical interventionsinaging, pages 1579–1587, 2017

    2017

  9. [17]

    Octid: Optical coherence tomography image database

    Peyman Gholami, Priyanka Roy, Mohana Kuppuswamy Parthasarathy, and Vasudevan Lakshminarayanan. Octid: Optical coherence tomography image database. Computers & Electrical Engineering, 81:106532, 2020

    2020

  10. [18]

    Kfwc: a knowledge-drivendeeplearningmodelforfine-grainedclassification of wet-amd.Computer Methods and Programs in Biomedicine, 229: 107312, 2023

    EHaihong,JiawenHe,TianyiHu,LifeiYuan,RuruZhang,Shengjuan Zhang, Yanhui Wang, Meina Song, and Lifei Wang. Kfwc: a knowledge-drivendeeplearningmodelforfine-grainedclassification of wet-amd.Computer Methods and Programs in Biomedicine, 229: 107312, 2023

    2023

  11. [19]

    Classifying neovascular age-related macular degeneration with a deep convolutional neural network based on optical coherence tomography images.ScientificReports, 12(1):2232, 2022

    Jinyoung Han, Seong Choi, Ji In Park, Joon Seo Hwang, Jeong Mo Han, Hak Jun Lee, Junseo Ko, Jeewoo Yoon, and Daniel Duck-Jin Hwang. Classifying neovascular age-related macular degeneration with a deep convolutional neural network based on optical coherence tomography images.Sc...

    2022

  12. [20]

    Aninterpretabletransformernetworkfortheretinaldisease classificationusingopticalcoherencetomography

    JingzhenHe,JunxiaWang,ZeyuHan,JunMa,ChongjingWang,and MengQi. Aninterpretabletransformernetworkfortheretinaldisease classificationusingopticalcoherencetomography. ScientificReports, 13(1):3637, 2023

    2023

  13. [21]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770– 778, 2016

    2016

  14. [22]

    Automated retinal disease classification using hybrid transformer model (svit) using optical coherence tomography images.NeuralComputingandApplications, pages 1–18, 2024

    GR Hemalakshmi, M Murugappan, Mohamed Yacin Sikkandar, S Sabarunisha Begum, and NB Prakash. Automated retinal disease classification using hybrid transformer model (svit) using optical coherence tomography images.NeuralComputingandApplications, pages 1–18, 2024

    2024

  15. [23]

    Automatic classification of retinal optical coherence tomography images with layer guided convolutional neural network

    LaifengHuang,XingxinHe,LeyuanFang,HosseinRabbani,andXi- angdong Chen. Automatic classification of retinal optical coherence tomography images with layer guided convolutional neural network. IEEESignal ProcessingLetters, 26(7):1026–1030, 2019

    2019

  16. [24]

    Optical coherence tomography to detect and manage retinal disease and glaucoma.American journal ofophthalmology, 137(1):156–169, 2004

    Glenn J Jaffe and Joseph Caprioli. Optical coherence tomography to detect and manage retinal disease and glaucoma.American journal ofophthalmology, 137(1):156–169, 2004

    2004

  17. [25]

    Subretinalfluidin macular edemasecondary to branchretinal vein occlusion.Scientific Reports, 14(1):13623, 2024

    DonghyunJee,SoyoungPark,andJin-wooKwon. Subretinalfluidin macular edemasecondary to branchretinal vein occlusion.Scientific Reports, 14(1):13623, 2024

    2024

  18. [26]

    Explainability of vision transformers: A comprehensive reviewandnewperspectives

    RojinaKashefi,LeiliBarekatain,MohammadSabokrou,andFatemeh Aghaeipoor. Explainability of vision transformers: A comprehensive reviewandnewperspectives. arXivpreprintarXiv:2311.06786,2023

    2023 arXiv

  19. [27]

    Identifying medical diagnoses and treatable diseases by image-based deep learning.cell, 172(5): 1122–1131, 2018

    Daniel S Kermany, Michael Goldbaum, Wenjia Cai, Carolina CS Valentim, Huiying Liang, Sally L Baxter, Alex McKeown, Ge Yang, Xiaokang Wu, Fangbing Yan, et al. Identifying medical diagnoses and treatable diseases by image-based deep learning.cell, 172(5): 1122–1131, 2018

    2018

  20. [28]

    Boost diagnostic performance in retinal disease classification utilizing deep ensemble classifiers based on oct.Multimedia Tools and Applications, pages 1–21, 2024

    Umer Sadiq Khan and Saif Ur Rehman Khan. Boost diagnostic performance in retinal disease classification utilizing deep ensemble classifiers based on oct.Multimedia Tools and Applications, pages 1–21, 2024

    2024

  21. [29]

    To- wards evaluating explanations of vision transformers for medical imaging

    Piotr Komorowski, Hubert Baniecki, and Przemyslaw Biecek. To- wards evaluating explanations of vision transformers for medical imaging. In Proceedings of the IEEE/CVF conference on computer visionand pattern recognition, pages 3726–3732, 2023

    2023

  22. [30]

    Octdl: Optical coherence tomography dataset for image-based deep learning methods.Scientific Data, 11(1):365, 2024

    Mikhail Kulyabin, Aleksei Zhdanov, Anastasia Nikiforova, An- drey Stepichev, Anna Kuznetsova, Mikhail Ronkin, Vasilii Borisov, Alexander Bogachev, Sergey Korotkich, Paul A Constable, et al. Octdl: Optical coherence tomography dataset for image-based deep learning methods.Scien...

    2024

  23. [31]

    Un- masking clever hans predictors and assessing what machines really learn

    Sebastian Lapuschkin, Stephan Wäldchen, Alexander Binder, Gré- goire Montavon, Wojciech Samek, and Klaus-Robert Müller. Un- masking clever hans predictors and assessing what machines really learn. Naturecommunications, 10(1):1096, 2019

    2019

  24. [32]

    Future-ai: International consensus guideline for trustworthy and deployable artificial intelligence in healthcare

    Karim Lekadir, Alejandro F Frangi, Antonio R Porras, Ben Glocker, CeliaCintas,CurtisPLanglotz,EvaWeicken,FolkertWAsselbergs, Fred Prior, Gary S Collins, et al. Future-ai: International consensus guideline for trustworthy and deployable artificial intelligence in healthcare. bm...

    2025

  25. [33]

    TatianaLikhomanenko,QiantongXu,GabrielSynnaeve,RonanCol- lobert,andAlexRogozhnikov.Cape:Encodingrelativepositionswith continuous augmented positional embeddings.Advances in Neural InformationProcessingSystems, 34:16079–16092, 2021

    2021

  26. [34]

    Focal loss for dense object detection

    T Lin. Focal loss for dense object detection. arXiv preprint arXiv:1708.02002, 2017

    2017 arXiv

  27. [35]

    Swintransformer:Hierarchicalvision transformerusingshiftedwindows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, StephenLin,andBainingGuo. Swintransformer:Hierarchicalvision transformerusingshiftedwindows. In ProceedingsoftheIEEE/CVF international conference on computer vision, pages 10012–10022, 2021

    2021

  28. [36]

    Decoupled weight decay regular- ization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regular- ization. arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  29. [37]

    Harder synthetic anomaliestoimproveooddetectioninmedicalimages

    Sergio Naval Marimont and Giacomo Tarroni. Harder synthetic anomaliestoimproveooddetectioninmedicalimages. arXivpreprint arXiv:2308.01412, 2023

    2023 arXiv

  30. [38]

    Using protopnet for interpretable alzheimer’s disease classification

    Sanaz Mohammadjafari, Mucahit Cevik, Mathusan Thanabal- asingam, and Ayse Basar. Using protopnet for interpretable alzheimer’s disease classification. InCanadian AI, 2021

    2021

  31. [39]

    Deep multimodal fusion of data with heterogeneous dimensionality via projective networks

    José Morano, Guilherme Aresta, Christoph Grechenig, Ursula Schmidt-Erfurth, and Hrvoje Bogunović. Deep multimodal fusion of data with heterogeneous dimensionality via projective networks. IEEEJournal of Biomedicaland Health Informatics, 2024

    2024

  32. [40]

    Trivialaugment: Tuning-free yet state-of-the-art data augmentation

    Samuel G Müller and Frank Hutter. Trivialaugment: Tuning-free yet state-of-the-art data augmentation. InProceedings of the IEEE/CVF international conferenceon computervision, pages 774–782, 2021

    2021

  33. [41]

    Neural pro- totype trees for interpretable fine-grained image recognition

    Meike Nauta, Ron Van Bree, and Christin Seifert. Neural pro- totype trees for interpretable fine-grained image recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14933–14943, 2021

    2021

  34. [42]

    Interpreting and correcting medical image classification with pip-net

    Meike Nauta, Johannes H Hegeman, Jeroen Geerdink, Jörg Schlöt- terer, Maurice van Keulen, and Christin Seifert. Interpreting and correcting medical image classification with pip-net. InEuropean ConferenceonArtificialIntelligence,pages198–215.Springer,2023

    2023

  35. [43]

    Pip-net: Patch-based intuitive prototypes for interpretable imageclassification

    Meike Nauta, Jörg Schlötterer, Maurice Van Keulen, and Christin Seifert. Pip-net: Patch-based intuitive prototypes for interpretable imageclassification. InProceedingsoftheIEEE/CVFConferenceon ComputerVisionand Pattern Recognition, pages 2744–2753, 2023

    2023

  36. [44]

    Marzieh Oghbaie, Teresa Araújo, Ursula Schmidt-Erfurth, and HrvojeBogunović.Vlfatrollout:Fullytransformer-basedclassifierfor retinal oct volumes.Computerized Medical Imaging and Graphics, page 102452, 2024

    2024

  37. [45]

    Dinov2: Learning robust vi- sual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Fran- cisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust vi- sual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  38. [46]

    Reza Rasti, Hossein Rabbani, Alireza Mehridehnavi, and Fedra Ha- jizadeh. Macular oct classification using a multi-scale convolutional neural network ensemble.IEEEtransactionsonmedicalimaging, 37 Oghbaie et al.:Preprint submitted to ElsevierPage 9 of 10 PiPViT (4):1024–1034, 2017

    2017

  39. [47]

    Stop explaining black box machine learning models forhighstakesdecisionsanduseinterpretablemodelsinstead

    Cynthia Rudin. Stop explaining black box machine learning models forhighstakesdecisionsanduseinterpretablemodelsinstead. Nature machineintelligence, 1(5):206–215, 2019

    2019

  40. [48]

    Imagenet large scale visual recognition challenge

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh,SeanMa,ZhihengHuang,AndrejKarpathy,AdityaKhosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115:211–252, 2015

    2015

  41. [49]

    arXivpreprintarXiv:2011.14340, 2020

    Dawid Rymarczyk, Łukasz Struski, Jacek Tabor, and Bartosz Zieliński.Protopshare:Prototypesharingforinterpretableimageclas- sificationandsimilaritydiscovery. arXivpreprintarXiv:2011.14340, 2020

    2011 arXiv

  42. [50]

    Protomil: Multiple instance learning with prototypical parts for whole-slide image clas- sification

    DawidRymarczyk,AdamPardyl,Jarosław Kraus,AnetaKaczyńska, Marek Skomorowski, and Bartosz Zieliński. Protomil: Multiple instance learning with prototypical parts for whole-slide image clas- sification. In Joint European Conference on Machine Learning and KnowledgeDiscoveryinData...

    2022

  43. [51]

    Spreading vectors for similarity search

    AlexandreSablayrolles,MatthijsDouze,CordeliaSchmid,andHervé Jégou. Spreading vectors for similarity search. arXiv preprint arXiv:1806.03198, 2018

    2018 arXiv

  44. [52]

    Transparency of deep neural networks for medical image analysis: A review of interpretability methods.Computers in biologyandmedicine, 140:105111, 2022

    Zohaib Salahuddin, Henry C Woodruff, Avishek Chatterjee, and Philippe Lambin. Transparency of deep neural networks for medical image analysis: A review of interpretability methods.Computers in biologyandmedicine, 140:105111, 2022

    2022

  45. [53]

    Towards explainable artificial intelligence

    Wojciech Samek and Klaus-Robert Müller. Towards explainable artificial intelligence. Explainable AI: interpreting, explaining and visualizingdeeplearning, pages 5–22, 2019

    2019

  46. [54]

    Exploitingepistemicuncertaintyofanatomy segmentation for anomaly detection in retinal oct.IEEE transactions onmedical imaging, 39(1):87–98, 2019

    PhilippSeeböck,JoséIgnacioOrlando,ThomasSchlegl,SebastianM Waldstein, Hrvoje Bogunović, Sophie Klimscha, Georg Langs, and UrsulaSchmidt-Erfurth. Exploitingepistemicuncertaintyofanatomy segmentation for anomaly detection in retinal oct.IEEE transactions onmedical imaging, 39(1)...

    2019

  47. [55]

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

    RamprasaathRSelvaraju,MichaelCogswell,AbhishekDas,Ramakr- ishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: visual explanations from deep networks via gradient-based localization. International journal of computervision, 128:336–359, 2020

    2020

  48. [56]

    Improving interpretabilityinmachinediagnosis:detectionofgeographicatrophy in oct scans.Ophthalmology Science, 1(3):100038, 2021

    Xiaoshuang Shi, Tiarnan DL Keenan, Qingyu Chen, Tharindu De Silva, Alisa T Thavikulwat, Geoffrey Broadhead, Sanjeeb Bhan- dari, Catherine Cukras, Emily Y Chew, and Zhiyong Lu. Improving interpretabilityinmachinediagnosis:detectionofgeographicatrophy in oct scans.Ophthalmology ...

    2021

  49. [57]

    Think positive: An interpretable neural network for image recognition.NeuralNetworks, 151:178–189, 2022

    Gurmail Singh. Think positive: An interpretable neural network for image recognition.NeuralNetworks, 151:178–189, 2022

    2022

  50. [58]

    Aninterpretabledeeplearning model for covid-19 detection with chest x-ray images.Ieee Access, 9:85198–85208, 2021

    GurmailSinghandKin-ChoongYow. Aninterpretabledeeplearning model for covid-19 detection with chest x-ray images.Ieee Access, 9:85198–85208, 2021

    2021

  51. [59]

    An interpretable and accurate deep-learning diagnosis framework modeled with fully and semi- supervised reciprocal learning

    Chong Wang, Yuanhong Chen, Fengbei Liu, Michael Elliott, ChunFungKwok,CarlosPena-Solorzano,HelenFrazer,DavisJames McCarthy, and Gustavo Carneiro. An interpretable and accurate deep-learning diagnosis framework modeled with fully and semi- supervised reciprocal learning. IEEE T...

    2023

  52. [60]

    Score-cam: Score- weighted visual explanations for convolutional neural networks

    Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score- weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognitionworkshops,...

    2020

  53. [61]

    Contrastive learning rivals masked image modeling in fine-tuning via feature distillation.arXiv preprint arXiv:2205.14141, 2022

    Yixuan Wei, Han Hu, Zhenda Xie, Zheng Zhang, Yue Cao, Jianmin Bao, Dong Chen, and Baining Guo. Contrastive learning rivals masked image modeling in fine-tuning via feature distillation.arXiv preprint arXiv:2205.14141, 2022

    2022 arXiv

  54. [62]

    Pytorch image models.https://github.com/ rwightman/pytorch-image-models, 2019

    Ross Wightman. Pytorch image models.https://github.com/ rwightman/pytorch-image-models, 2019

    2019

  55. [63]

    Wan Ling Wong, Xinyi Su, Xiang Li, Chui Ming G Cheung, Ronald Klein, Ching-Yu Cheng, and Tien Yin Wong. Global prevalence of age-related macular degeneration and disease burden projection for 2020 and 2040: a systematic review and meta-analysis.The Lancet GlobalHealth, 2(2):e1...

    2020

  56. [64]

    Protopformer: Concentrating on prototypical parts in vision transformers for interpretable image recognition

    Mengqi Xue, Qihan Huang, Haofei Zhang, Lechao Cheng, Jie Song, Minghui Wu, and Mingli Song. Protopformer: Concentrating on prototypical parts in vision transformers for interpretable image recognition. arXiv preprint arXiv:2208.10431, 2022

    2022

  57. [65]

    Survey on explainable ai: From approaches, limitations and applications aspects

    Wenli Yang, Yuchen Wei, Hanyu Wei, Yanyu Chen, Guan Huang, Xiang Li, Renjie Li, Naimeng Yao, Xinyi Wang, Xiaotong Gu, et al. Survey on explainable ai: From approaches, limitations and applications aspects. Human-Centric Intelligent Systems, 3(3):161– 188, 2023

    2023

  58. [66]

    xx”, non-prototypical models with “–

    Yukun Zhou, Mark A Chia, Siegfried K Wagner, Murat S Ayhan, Dominic J Williamson, Robbert R Struyven, Timing Liu, Moucheng Xu, Mateo G Lozano, Peter Woodward-Court, et al. A foundation modelforgeneralizablediseasedetectionfromretinalimages. Nature, 622(7981):156–163, 2023. Ogh...

    2023

Pith tools

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