Pith. sign in

REVIEW 2 major objections 5 minor 33 references

endoExplain: A reproducible protocol for auditing score-localisation discordance in colonoscopy image classifiers

T0 review · 2 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read A high classifier score and a plausible heatmap do not by themselves show where a colonoscopy model is looking, and this paper demonstrates that with a reproducible audit protocol.

desk verdict A thorough, honest protocol paper whose central point—high scores don't imply localisation—is well supported; the mask-reference caveat is real but doesn't sink it. read the letter →

arxiv 2607.19372 v2 pith:ZZXKUOYE submitted 2026-06-16 eess.IV q-bio.QM

Roberto Alcaraz Machado , Ana Lucía Rodríguez Blanco This is my paper
classification eess.IVq-bio.QM
keywords colonoscopyexplainableAIclassactivationmappingcalibrationlocalisationaudittemperaturescalingperturbationresponsereproducibility
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces endoExplain, a reproducible protocol for auditing whether a colonoscopy image classifier's confidence score and its class-activation map actually point at the same evidence. It does not propose a new detector or explanation method; it coordinates content-hash data separation, validation-only temperature scaling, paired multi-method evaluation on identical image-mask pairs, and separate reporting of spatial overlap and perturbation response. On 172 reserved images with temperature-scaled score at least 0.90, the rate at which a CAM's peak fell outside the lesion ranged from 4.1% to 62.2% depending on which CAM was used, and rankings were not stable across architectures, seeds, thresholds, or external datasets. Deletion response also failed to separate methods: random deletion produced a larger mean logit drop than any CAM in the high-score population. The authors conclude that neither a high score nor a visually persuasive heatmap is evidence of lesion localisation or model reasoning, and they expose a source-category confounder: 149 of 155 test positives were one polyp subtype, so aggregate performance does not generalise to ordinary polyp detection.

What carries the argument

The central object is the paired, score-conditioned localisation audit: for each image, a temperature-scaled positive score s_i, a CAM map H_i^m, and a binary lesion mask M_i are locked together, and the primary endpoint is a pointing-game indicator F_i^m(c) = 1(s_i >= c) * 1[argmax H_i^m not in M_i], reported conditionally on s_i >= c. This indicator, plus a separate top-20% overlap measure and a top-20% deletion logit-drop with a random-deletion control, is what carries the argument: it makes calibration, spatial agreement, perturbation response, and transfer separately inspectable on identical image-mask pairs.

What would settle it

A decisive test would be to occlude or inpaint everything outside the lesion mask and measure the target logit on the same 172 high-score images: if the model still assigns scores at or above 0.90 when only the mask contents remain, then the mask is the evidence location and the reported discordance rates are meaningful; if scores collapse when context is removed, then the mask is not the right ground truth and the 4.1-62.2% range is an artifact of the reference, not of the CAMs.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that score and heatmap are not interchangeable evidence. After fitting a scalar temperature on validation data only, 172 test images had scaled score at least 0.90; on those identical images, Grad-CAM++ failed to peak inside the annotated lesion 62.2% of the time while HiResCAM failed 5.8% and Eigen-CAM 4.1%. The discordance persisted across score thresholds, architectures, and training seeds, although the ranking of methods was not universal. Spatial alignment and deletion response disagreed within the same maps: in the full 1,000-image audit, Eigen-CAM had the highest mean overlap but a negative mean logit drop, while random deletion produ

Load-bearing premise

The protocol's load-bearing premise is that the binary lesion mask is the correct spatial reference for 'where the evidence lies'; the paper itself states in its limitations that the mask does not encode every context feature legitimately used by a classifier, so if context outside the mask carries legitimate evidence, then 'peak-outside-lesion' does not necessarily mean the model is looking at the wrong place.

Editorial extensions

If this is right

  • If the central claim is right, a high temperature-scaled score threshold cannot be used to claim the model has found the lesion; 4.1-62.2% of such images have a CAM peak outside the mask depending on the map.
  • If the central claim is right, heatmap displays should show the CAM method, score-calibration status, and threshold, because CAM rankings are not universal across architectures, seeds, and datasets.
  • If the central claim is right, a positive deletion response is not evidence of attribution specificity, since an exact-area random-deletion control produced a larger mean logit drop than all CAMs in the high-score population.
  • If the central claim is right, external deployment of a frozen model requires dataset-specific spatial auditing, since mean overlap changed across the three external mask cohorts and a centre baseline reached similar overlap.
  • If the central claim is right, classifier performance claims need source-category composition, because aggregate AUPRC was dominated by dyed-lifted polyps and standard-polyp sensitivity was 0.333 on six test images.

Reading between the lines

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

  • Editorial inference, not in the paper: a natural testable extension would be to repeat the audit on a model trained with context-erased inputs, such as inpainting the lesion, to see whether outside-mask peaks carry causal evidence; the paper's own limitation section says the mask does not encode every legitimate context feature, which is exactly the assumption this test would probe.
  • Editorial inference, not in the paper: the protocol could be turned into a runtime monitor for deployed computer-aided detection, flagging frames where score exceeds a threshold but the CAM peak falls outside the polyp boundary, and logging those frames for human review.
  • Editorial inference, not in the paper: applying the same paired audit to transformer-based attention maps would reveal whether attention-as-explanation inherits the same method-dependence, since attention is not gradient-weighted the way CAMs are.
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

2 major / 5 minor

Summary. The paper introduces endoExplain, a reproducible protocol for auditing score–localisation discordance in colonoscopy image classifiers. Using content-hash separation of HyperKvasir development and masked images, the authors train three architectures with three seeds each, fit temperature scaling on validation data only, and evaluate five CAM methods plus random and centre baselines on identical image–mask pairs. The main empirical result is that among 172 reserved standard-polyp images with temperature-scaled score ≥0.90, peak-outside-lesion rates range from 4.1% (Eigen-CAM) to 62.2% (Grad-CAM++), while a random-deletion control also produces positive logit drops. The paper reports robustness to thresholds, architecture/seed, lesion size and centrality, and frozen external transfer to three mask cohorts, and it includes a source-category audit showing that 149/155 test positives are dyed-lifted polyps. The authors conclude that classifier score, CAM localisation, and perturbation response are separate constructs, and caution against treating a high score or visually plausible CAM as evidence of lesion localisation or model reasoning.

Significance. If the results hold, the paper makes a valuable methodological contribution: it provides a concrete, code-linked template for separating calibration, spatial alignment, perturbation response, and transfer in explainability audits. The inclusion of content-defined data roles, validation-only model selection and temperature fitting, paired per-image CAM comparison, spatial baselines, random-deletion controls, multiple seeds, and a source-category audit is unusually rigorous for this literature. The finding that method choice changes spatial interpretation on identical inputs — e.g., Grad-CAM failure 50.6% versus HiResCAM 5.8% on the same 172 images — is a strong, falsifiable demonstration that CAMs are not interchangeable. The paper's explicit recognition of its limitations, including mask incompleteness and the dyed-lifted-polyp confound, further supports the credibility of the protocol. The exact discordance rates, however, are conditional on the binary lesion mask as the spatial reference, and this assumption needs additional robustness analysis before the strongest wording is fully supported.

major comments (2)
  1. [§5.4, Eq. (2), Table 4] The primary endpoint treats the HyperKvasir binary lesion mask as the complete spatial reference, yet the paper itself states in Section 5.4 that the mask does not encode every context feature legitimately used by a classifier. As a result, the headline 4.1–62.2% peak-outside-lesion range is conditional on mask validity. Please add a mask-dilation sensitivity analysis (e.g., 0/5/10/20 pixels) and report whether the peak-outside rates and method ordering change. If the rates are materially affected, the conclusion that 'a high score is not evidence of lesion localisation' should be tempered to reflect that the reference itself is incomplete.
  2. [§4.1, Table 2, §4.2] The high-score audited pool is distribution-shifted: the model was trained almost entirely on dyed-lifted polyps (702/721 train positives), while the 1,000 spatial-audit images are standard polyps. The test set contains only 6 standard polyps, and the model's sensitivity on them is 0.333. Thus the 172 images with score ≥0.90 are a highly selected subset of standard polyp images, potentially enriched for cases where the model relies on procedural context rather than lesion appearance. Please quantify the sensitivity of the discordance rates to this compositional shift — for example, by reporting score distributions for standard vs. dyed-lifted positives or by repeating the primary analysis on strongly diluted masks — and adjust the generalization claims accordingly.
minor comments (5)
  1. [§3.3] The tie-breaking rule for constant maps and boundary ties is reported, but the multiplicative permutation used for tied top-area pixels is not specified. Please provide the exact seed and permutation construction in the reproducibility appendix.
  2. [§4.1, Table 3] Temperature scaling improves ECE and NLL but slightly worsens Brier score. This is correctly reported; it would be helpful to add a sentence explaining that a single scalar calibration can trade off different proper scores, so ECE alone should not be read as evidence of overall improvement.
  3. [§4.5, Figure 5] For the ETIS cohort, only 16 images are eligible at the 0.90 threshold. The text says the subset is too small for stable rankings, but the figure still shows the full bar set. Please add a visual or textual marker that the ETIS estimates have very wide intervals or omit the conditional bars for ETIS.
  4. [§5.4] The limitations paragraph notes that patient identifiers were unavailable and image bootstrap intervals may be too narrow. This is appropriately candid, but the paper would be strengthened by also stating explicitly how many images might share a procedure or sequence source, if this is known from the dataset metadata.
  5. [§2, Related work] Reference [7] (HiResCAM) is a preprint; please add a citation to the peer-reviewed version if one exists, or state that the preprint was the definitive source used.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the score-localisation audit uses pre-declared, validation-only fitted calibration and a hash-disjoint held-out spatial pool; no target outcome is fitted and then predicted.

full rationale

The paper's derivation chain is self-contained and non-circular. Training, checkpoint selection (validation AUPRC), and temperature fitting (validation NLL only) are completed before the spatial audit, and the 1,000 masked images were hash-separated before training. The primary outcome, Eq. (2), is a descriptive conditional failure indicator: F_im(c)=1(s_i>=c)1[argmax H_im notin M_i]. The reported 4.1%-62.2% peak-outside-lesion rates are proportions within the 172 images meeting the predeclared threshold; they are not used to fit any parameter, and the temperature parameter is not fitted to the CAM or mask outcomes. The paper explicitly reports that temperature scaling changed test ECE from 0.0167 to 0.0115 and also worsened Brier score, showing the result is not a selected fit. The acknowledged limitation in Section 5.4 that 'The lesion mask is treated as a spatial reference but does not encode every context feature legitimately used by a classifier' concerns the validity of the mask as spatial ground truth, not circularity: no conclusion is derived by assuming the conclusion. The source-category confounding (149/155 test positives dyed-lifted polyps) is likewise an interpretability and generalisability limitation that the paper itself surfaces. No load-bearing self-citation is present; the HiResCAM faithfulness property is cited from external prior work and explicitly hedged as not making a CAM a lesion mask. The random-deletion and centre baselines provide independent controls. Therefore no circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central empirical findings rest on two fitted quantities (temperature T and the validation-selected checkpoint) and on domain assumptions about content separation, mask validity, and metric semantics. The paper is transparent about most of these; no invented entities are introduced.

free parameters (2)
  • Temperature scaling T = 0.829
    Fitted on validation split by minimising binary NLL; defines the temperature-scaled score threshold (>=0.90) that selects the 172-image primary high-score population. If refit on test or spatial data, the rates would change.
  • Primary checkpoint selection = EfficientNet-B0 seed 2089
    Selected among three EfficientNet-B0 seeds by validation AUPRC (0.983). The RQ1 table uses this single checkpoint; architecture and seed sensitivity are reported separately, but the headline 4.1-62.2% range is checkpoint-specific.
assumptions (4)
  • domain assumption Content separation by SHA-256 byte identity and exact 256-bit average-hash fingerprint guarantees no training/evaluation leakage into the spatial audit.
    Central to the claim that the 1,000 masked images are unseen. SHA-256 is exact, but the average-hash rule only catches identical fingerprints (Hamming distance 0), not near-duplicates. Location: Section 3.1.
  • domain assumption The binary lesion mask is an adequate reference for what the classifier should localise; disagreement with the mask is interpreted as score-localisation discordance.
    The paper itself notes in Section 5.4 that the mask does not encode every context feature legitimately used by a classifier, so CAM peaks outside the mask may reflect context use rather than failure. If the mask is wrong for this purpose, the headline discordance interpretation weakens.
  • domain assumption Pointing-game peak, top-20% IoU, and deletion logit-drop are meaningful and separable properties of a CAM.
    The protocol's outcomes rely on these metrics; the paper acknowledges they are not causal explanation metrics (Section 5.4) and that mean-value deletion is out of distribution (Section 3.4).
  • ad hoc to paper Combining polyps and dyed-lifted-polyps into a binary polyp_family target yields a meaningful audit construct.
    This collapsed target drives training and the score conditioning; the source-category audit shows the model mostly separates dyed-lifted polyps, so generalising to standard polyps is limited. Location: Section 3.1 and Table 2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of endoExplain: A reproducible protocol for auditing score-localisation discordance in colonoscopy image classifiers." pith.science (2026). https://pith.science/paper/ZZXKUOYE

@misc{pith2026260719372,
  author       = {Pith},
  title        = {Pith review of: endoExplain: A reproducible protocol for auditing score-localisation discordance in colonoscopy image classifiers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZZXKUOYE}},
  note         = {Machine review of arXiv:2607.19372}
}
read the original abstract

Background and objective: A high classifier score and a plausible class-activation map (CAM) are often presented together, although neither establishes that the other is reliable. We introduce endoExplain as a reproducible protocol for auditing score-localisation discordance rather than as a new detector or explanation algorithm. Methods: Content hashing separated HyperKvasir development images from 1,000 masked images before training. EfficientNet-B0, ResNet-34 and ConvNeXt-Tiny were trained with three seeds each. Scores were temperature scaled using validation data only. Grad-CAM, Grad-CAM++, XGrad-CAM, HiResCAM and Eigen-CAM were evaluated on identical image-mask pairs, alongside random and centre baselines. Outcomes combined peak localisation, overlap, a top-20% deletion response, score-threshold sensitivity and adjustment for lesion size and centrality. The selected checkpoint was transferred without retraining to three external mask cohorts. Results: Temperature scaling reduced test expected calibration error from 0.0167 to 0.0115. Among 172 reserved images with scaled score at least 0.90, peak-outside-lesion rates ranged from 4.1% to 62.2% across CAMs. Method dependence remained evident across architectures and seeds, although method rankings were not universal. Spatial alignment and deletion response were not interchangeable. A random-deletion control also produced positive logit drops, limiting specificity claims based on deletion alone. External positive-mask results were dataset dependent. A source-category audit also exposed that 149/155 test positives were dyed-lifted polyps, materially bounding classifier claims. Conclusions: endoExplain makes calibration, spatial agreement, perturbation response and transfer separately inspectable. The results caution against using a score or visually persuasive CAM as evidence of lesion localisation or model reasoning.

Figures

Figures reproduced from arXiv: 2607.19372 by the authors.

Figure 1
Figure 1. endoExplain workflow. Content-defined roles precede model development. Calibration uses validation data only. Localisation and perturbation outcomes are calculated on paired maps, followed by robustness analyses and transfer without retraining. HyperKvasir supplied labelled lower-gastrointestinal images and a segmented polyp subset [14]. SHA-256 comparison established that all 1,000 segmented images were byte-identi… view at source ↗
Figure 2
Figure 2. Calibration and score-threshold sensitivity. (A) Reliability of raw and validation-scaled scores on the held-out classifier test split. (B) Peak-outside-lesion rate on the reserved mask pool as the scaled-score threshold changes; labels give the number of eligible images. Eigen-CAM is class-agnostic and marked with an asterisk. 4.2 RQ1: high-score localisation discordance For the primary checkpoint, 172 of 1,000 sta… view at source ↗
Figure 3
Figure 3. Rule-selected qualitative evidence. Row A is CAM-discordant and row B is five-CAM consensus. Columns retain the original image and explicit lesion mask. All CAM panels use identical contrast conventions (greyscale anatomy, inferno activation, cyan mask, white peak). Rendered metrics are checked against the paired export within a 0.002 IoU tolerance. Eigen-CAM is class-agnostic. Source images: HyperKvasir, Borgli et … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Architecture–seed robustness across all nine checkpoints. Points are architecture means and whiskers are across-seed ranges. (A) Peak failure on each checkpoint’s own score-defined subset (𝑛 = 38–356). (B) Peak failure on the fixed 172 images selected by the primary ch…
Figure 5
Figure 5. Figure 5: Frozen visible-mask transfer. Bars show mean top-20% CAM–mask IoU with 95% image￾bootstrap intervals for the selected checkpoint. The three external cohorts contain no compatible negative controls; 125 mask-empty PolypGen frames were excluded from localisation. Dashed …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 7 canonical work pages

  1. [1]

    Development and validation of a deep-learning algorithm for the detection of polyps during colonoscopy

    Pu Wang, Tyler M Berzin, Jeremy Romek Glissen Brown, Shishira Bharadwaj, Aymeric Becq, Xun Xiao, Peixi Liu, Liangping Li, Yan Song, Difei Zhang, et al. Development and validation of a deep-learning algorithm for the detection of polyps during colonoscopy. Nature Biomedical Engineering, 2(10):741–748, 2018. doi: 10.1038/s41551-018-0301-3

    work page doi:10.1038/s41551-018-0301-3 2018

  2. [2]

    Diagnostic accuracy of artificial intelligence in endoscopy: Umbrella review.JMIR Medical Informatics, 12:e56361, 2024

    Bowen Zha, Angshu Cai, and Guiqi Wang. Diagnostic accuracy of artificial intelligence in endoscopy: Umbrella review.JMIR Medical Informatics, 12:e56361, 2024. doi: 10.2196/ 56361

    2024

  3. [3]

    A prospective multicenter randomized controlled trial on artificial intelligence assisted colonoscopy for enhanced polyp detection.Scientific Reports, 14(1):25453, 2024

    Dong Kyun Park, Eui Joo Kim, Jong Pil Im, Hyun Lim, Yun Jeong Lim, Jeong-Sik Byeon, Kyoung Oh Kim, Jun-Won Chung, Yoon Jae Kim, et al. A prospective multicenter randomized controlled trial on artificial intelligence assisted colonoscopy for enhanced polyp detection.Scientific Reports, 14(1):25453, 2024. doi: 10.1038/s41598-024-77079-1

    work page doi:10.1038/s41598-024-77079-1 2024

  4. [4]

    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. InProceedings of the IEEE International Conference on Computer Vision (ICCV), pages 618–626, 2017. doi: 10.1109/ICCV.2017.74

    doi:10.1109/iccv.2017.74 2017

  5. [5]

    Grad-cam++: Generalized gradient-based visual explanations for deep convolutional net- works

    Aditya Chattopadhay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional net- works. InIEEE Winter Conference on Applications of Computer Vision (WACV), pages 839–847, 2018. doi: 10.1109/WACV.2018.00097

    arXiv 2018

  6. [6]

    Axiom- based Grad-CAM: Towards accurate visualization and explanation of CNNs

    Ruigang Fu, Qingyong Hu, Xiaohu Dong, Yulan Guo, Yinghui Gao, and Biao Li. Axiom- based Grad-CAM: Towards accurate visualization and explanation of CNNs. InBritish Machine Vision Conference, 2020

    2020

  7. [7]

    Use HiResCAM instead of Grad-CAM for faithful explanations of convolutional neural networks, 2021

    Rachel Lea Draelos and Lawrence Carin. Use HiResCAM instead of Grad-CAM for faithful explanations of convolutional neural networks, 2021

    2021

  8. [8]

    Sanity checks for saliency maps

    Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. InAdvances in Neural Information Processing Systems (NeurIPS), 2018. 15

    2018

Show all 33 references
  1. [9]

    Evaluating the visualization of what a deep neural network has learned

    Wojciech Samek, Alexander Binder, Grégoire Montavon, Sebastian Lapuschkin, and Klaus- Robert Müller. Evaluating the visualization of what a deep neural network has learned. IEEE Transactions on Neural Networks and Learning Systems, 28(11):2660–2673, 2017. doi: 10.1109/TNNLS.20...

    2017

  2. [10]

    A benchmark for interpretability methods in deep neural networks

    Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. A benchmark for interpretability methods in deep neural networks. InAdvances in Neural Information Processing Systems, volume 32, 2019

    2019

  3. [11]

    A consistent and efficient evaluation strategy for attribution methods

    Yao Rong, Tobias Leemann, Vadim Borisov, Gjergji Kasneci, and Enkelejda Kasneci. A consistent and efficient evaluation strategy for attribution methods. InProceedings of the 39th International Conference on Machine Learning, volume 162 ofProceedings of Machine Learning Researc...

    2022

  4. [12]

    Uncertainty and inter- pretability in convolutional neural networks for semantic segmentation of colorectal polyps

    Kristoffer Wickstrøm, Michael Kampffmeyer, and Robert Jenssen. Uncertainty and inter- pretability in convolutional neural networks for semantic segmentation of colorectal polyps. Medical Image Analysis, 60:101619, 2020. doi: 10.1016/j.media.2019.101619

    2020

  5. [13]

    Assessing the trust- worthiness of saliency maps for localizing abnormalities in medical imaging.Radiology: Artificial Intelligence, 3(6):e200267, 2021

    Nishanth Arun, Nathan Gaw, Praveer Singh, Ken Chang, Mehak Aggarwal, Bryan Chen, Katharina Hoebel, Sharut Gupta, Jay Patel, Mishka Gidwani, et al. Assessing the trust- worthiness of saliency maps for localizing abnormalities in medical imaging.Radiology: Artificial Intelligenc...

    2021 doi

  6. [14]

    Hyperkvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy.Scientific Data, 7(1):283, 2020

    Hanna Borgli, Vajira Thambawita, Pia H Smedsrud, Steven Hicks, Debesh Jha, Sigrun L Eskeland, Kristin Ranheim Randel, Konstantin Pogorelov, Mathias Lux, Duc Tien Dang Nguyen, et al. Hyperkvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy...

    2020 doi

  7. [15]

    Toward embedded detection of polyps in wce images for early diagnosis of colorectal cancer.In- ternational Journal of Computer Assisted Radiology and Surgery, 9(2):283–293, 2014

    Juan Silva, Aymeric Histace, Olivier Romain, Xavier Dray, and Bertrand Granado. Toward embedded detection of polyps in wce images for early diagnosis of colorectal cancer.In- ternational Journal of Computer Assisted Radiology and Surgery, 9(2):283–293, 2014. doi: 10.1007/s1154...

    2014 doi

  8. [16]

    Salem, Dominique Lamarque, Christian Daul, Michael A

    Sharib Ali, Debesh Jha, Noha Ghatwary, Stefano Realdon, Renato Cannizzaro, Osama E. Salem, Dominique Lamarque, Christian Daul, Michael A. Riegler, Kim V. Anonsen, Andreas Petlund, Pål Halvorsen, Jens Rittscher, Thomas de Lange, and James E. East. A multi- centre polyp detectio...

    2023 doi

  9. [17]

    Debesh Jha, Nikhil Kumar Tomar, Vanshali Sharma, Quoc-Huy Trinh, Koushik Biswas, Hongyi Pan, Ritika K. Jha, Gorkem Durak, Alexander Hann, Jonas Varkey, Hang Viet Dao, Long Van Dao, Binh Phuc Nguyen, Nikolaos Papachrysos, Brandon Rieders, Peter Thelin Schmidt, Enrik Geissler, T...

    2024 doi arXiv

  10. [18]

    Eigen-cam: Class activation map using principal components

    Mohammed Bany Muhammad and Mohammed Yeasin. Eigen-cam: Class activation map using principal components. InInternational Joint Conference on Neural Networks (IJCNN), pages 1–7, 2020. doi: 10.1109/IJCNN48605.2020.9206626

    2020

  11. [19]

    Anna Hedström, Leander Weber, Daniel Krakowczyk, Dilyara Bareeva, Franz Motzkus, Wojciech Samek, Sebastian Lapuschkin, and Marina M.-C. Höhne. Quantus: An explainable AI toolkit for responsible evaluation of neural network explanations and beyond.Journal of Machine Learning Re...

    2023

  12. [20]

    RISE:Randomizedinputsamplingforexplanation of black-box models

    VitaliPetsiuk, AbirDas, andKateSaenko. RISE:Randomizedinputsamplingforexplanation of black-box models. InBritish Machine Vision Conference, 2018

    2018

  13. [21]

    Storås, Maximilian Dreyer, Frederik Pahde, Sebastian Lapuschkin, Wojciech Samek, Pål Halvorsen, Thomas de Lange, Yuichi Mori, Alexander Hann, Tyler M

    Andrea M. Storås, Maximilian Dreyer, Frederik Pahde, Sebastian Lapuschkin, Wojciech Samek, Pål Halvorsen, Thomas de Lange, Yuichi Mori, Alexander Hann, Tyler M. Berzin, Sravanthi Parasa, and Michael A. Riegler. Exploring the clinical value of concept-based AI explanations in g...

    2025 doi

  14. [22]

    Lago, Ghada Zamzmi, Brandon Eich, and Jana G

    Miguel A. Lago, Ghada Zamzmi, Brandon Eich, and Jana G. Delfino. Evaluating explainabil- ity: A framework for systematic assessment of explainable AI features in medical imaging. Bioengineering, 13(1):111, 2026. doi: 10.3390/bioengineering13010111

    2026 doi

  15. [23]

    artificial intelligence in endoscopy

    Yuchi Mori and Masashi Misawa. Quality assessment in endoscopy “artificial intelligence in endoscopy”.Best Practice & Research Clinical Gastroenterology, 76:102006, 2025. doi: 10.1016/j.bpg.2025.102006

    2025

  16. [24]

    LDPolypVideo benchmark: A large-scale colonoscopy video dataset of diverse polyps

    Yiting Ma, Xuejin Chen, Kai Cheng, Yang Li, and Bin Sun. LDPolypVideo benchmark: A large-scale colonoscopy video dataset of diverse polyps. InMedical Image Computing and Computer Assisted Intervention – MICCAI 2021, pages 387–396, 2021. doi: 10.1007/ 978-3-030-87240-3_37

    2021

  17. [25]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. InInternational Conference on Machine Learning (ICML), pages 6105–6114, 2019

    2019

  18. [26]

    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 (CVPR), pages 770–778, 2016. doi: 10.1109/CVPR.2016.90

    2016 doi

  19. [27]

    A convnet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11976–11986, 2022. doi: 10.1109/CVPR52688.2022.01167

    2022

  20. [28]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. InAdvances in Neural Information Processing Systems (N...

    2019

  21. [29]

    Torchvision the machine-vision package of torch

    Sébastien Marcel and Yann Rodriguez. Torchvision the machine-vision package of torch. In Proceedings of the 18th ACM International Conference on Multimedia, pages 1485–1488,

  22. [30]

    Albumentations: Fast and flexible image augmentations

    Alexander Buslaev, Vladimir I Iglovikov, Eugene Khvedchenya, Alex Parinov, Mikhail Druzhinin, and Alexandr A Kalinin. Albumentations: Fast and flexible image augmentations. Information, 11(2):125, 2020. doi: 10.3390/info11020125

    2020 doi

  23. [31]

    Mixed precision training

    Paulius Micikevicius, Sharan Narang, Jonah Alben, Gregory Diamos, Erich Elsen, David Garcia, Boris Ginsburg, Michael Houston, Oleksii Kuchaiev, Ganesh Venkatesh, et al. Mixed precision training. InInternational Conference on Learning Representations (ICLR), 2018

    2018

  24. [32]

    Weinberger

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. InProceedings of the 34th International Conference on Machine Learning, volume 70 ofProceedings of Machine Learning Research, pages 1321–1330, 2017. 18

    2017

  25. [2010]

    doi: 10.1145/1873951.1874254

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.