Pith. sign in

REVIEW 4 major objections 5 minor 58 references

Explanation stability is an interaction between the model and the attribution method, not an intrinsic property of the model alone.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 20:18 UTC pith:SKCDMMK6

load-bearing objection The position is reasonable, the numbers are not: the paper's own tables contradict the headline reversal, so the empirical core needs a major rework before the argument can be trusted. the 4 major comments →

arxiv 2607.16652 v1 pith:SKCDMMK6 submitted 2026-07-18 cs.CV cs.AI

Position: Explanation Stability Is a Property of the Model Method Pair, Not the Model

Kabilan Elangovan , Daniel Ting This is my paper
classification cs.CV cs.AI
keywords explanation stabilityattribution methodsGrad-CAM++LayerCAMsemantic drifttransfer learningXAI evaluationmedical imaging
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This position paper argues that claims about an AI model's explanation stability are scientifically invalid unless the attribution method is specified, just as statistical significance claims require specifying the test statistic. The evidence: three chest X-ray classifiers with nearly identical accuracy (>99% AUC) produce opposite stability rankings depending on which attribution method visualizes them. LayerCAM ranks InceptionV3 most stable, while Grad-CAM++ ranks DenseNet201 first and cuts InceptionV3's stability score by 17.3%. Both rankings are valid within their computational frameworks, so the paper concludes that stability is an emergent property of the model–method pair. The stakes are practical: single-method stability studies and regulatory submissions that omit the attribution operator can create illusory safety assurances.

Core claim

Using a transfer-learning to fine-tuning experiment on a five-class chest X-ray task, the paper shows that the stability of a model's saliency maps depends on the attribution operator chosen. With LayerCAM, InceptionV3 appears most stable (IoU 0.777), ahead of DenseNet201 (0.699) and ResNet50V2 (0.519). With Grad-CAM++, the ranking reverses: DenseNet201 leads (0.690), InceptionV3 drops to 0.643 (a 17.3% cut), and ResNet50V2 falls to 0.383. The reversal persists per-pathology (InceptionV3's COVID-19 IoU falls from 0.720 to 0.404) and occurs while every model keeps AUC above 0.99, so predictive accuracy cannot predict explanation behavior. DenseNet201 is the only architecture whose stability i

What carries the argument

The demonstration rests on a semantic-drift measurement: for each model, saliency maps are computed at epoch 8 (transfer-learning plateau) and epoch 19 (fine-tuning convergence), thresholded into binary masks, and compared using IoU — intersection-over-union of the masked regions. Samples are restricted to images all three models classify correctly in both phases (true-positive filtering, 2,430 of 3,354 test images), so measured drift reflects explanation evolution rather than accuracy changes. The two attribution methods are chosen for distinct mathematical objectives: LayerCAM uses pixel-wise gradients for fine-grained localization, while Grad-CAM++ uses higher-order, globally pooled gradi

Load-bearing premise

The central reversal depends on LayerCAM genuinely finding InceptionV3 more stable than DenseNet201, yet that specific comparison is statistically negligible (effect size d ≈ 0.10) and could be noise; the general conclusion also assumes two gradient-based explainers stand in for attribution methods as a whole.

What would settle it

Re-run the same true-positive-filtered protocol with a third attribution method from a different paradigm (for example, a path-based method such as Integrated Gradients) on a held-out chest X-ray dataset: if the three architectures keep a single stability ranking under every method, the reversal claim collapses. A cheaper check is to test whether InceptionV3's small LayerCAM lead over DenseNet201 (effect size ≈ 0.10) replicates on a larger sample; if it does not, the LayerCAM ordering is noise.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Researchers cannot claim general explanation stability from a single attribution method; findings must be validated across methods with different mathematical foundations or explicitly scoped to one method's computational objective.
  • Regulatory submissions for medical AI devices should specify the attribution operator used, and changing methods should trigger re-validation, mirroring algorithm-locking requirements.
  • Architecture selection for explanation-critical deployment should weigh cross-method robustness, not accuracy alone; the paper identifies dense connectivity as a design principle for method-robust explanations.
  • XAI benchmarks should report cross-method variance as a first-class dimension and penalize models whose stability rankings are method-sensitive.
  • Explanation stability and classification accuracy behave as orthogonal dimensions and must be evaluated independently.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The argument implies an operational audit rule: if a deployment pipeline can swap its explanation tool without touching the model, any stability guarantee should be re-verified under the new tool — a rule regulators could implement directly.
  • Given the paper's own citation of a no-free-lunch-style limit on explanations, the ranking reversal likely persists for perturbation-, path-, and attention-based methods; a third attribution paradigm would be the natural stress test.
  • The architectural claim is testable as a screening principle: models with broad feature sharing (dense or globally integrating connectivity) should show lower cross-method variance than parallel-pathway or shortcut-heavy models across a larger architecture zoo.
  • A collateral reading of the literature follows: prior single-method stability results are uninterpretable as model properties and may need re-ranking once the attribution operator is taken into account.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. This position paper argues that claims about explanation stability are scientifically invalid unless scoped to, or validated across, attribution methods. The authors support this with controlled chest X-ray experiments in which three ImageNet-pretrained CNNs (DenseNet201, ResNet50V2, InceptionV3) are trained in two phases (frozen-backbone transfer learning, then fine-tuning). Explanation drift between epochs 8 and 19 is measured with LayerCAM and Grad-CAM++ on a true-positive-filtered test set, using overlap IoU as the primary metric. The paper reports that stability rankings reverse across the two attribution methods, with LayerCAM favoring InceptionV3 and Grad-CAM++ favoring DenseNet201, and then draws normative conclusions for XAI evaluation practice, regulatory submissions, architecture selection, and benchmark design.

Significance. If the empirical demonstration were reliable, the paper would make an important and timely point. The recommendation that stability claims name the attribution operator parallels the requirement to name the test statistic, and it has direct regulatory relevance given recent FDA guidance. The position also aligns with existing published evidence of method-dependent attribution rankings, such as Duan et al. and MetaQuantus, and with no-free-lunch impossibility results cited in the paper. The true-positive filtering design is a thoughtful attempt to separate explanation evolution from prediction-error correction. However, the paper's own empirical core is currently undermined by numerous internal numerical contradictions. The reader cannot determine which table or figure is authoritative, and the claimed LayerCAM ranking advantage of InceptionV3 over DenseNet201 is contradicted by the paper's own Appendix B effect-size analysis. The conceptual position may survive on the strength of prior literature, but the controlled experiments as reported do not currently establish the paper's central demonstration.

major comments (4)
  1. [§3.3, Table 2 vs Appendix B (Table 4)] The load-bearing claim that LayerCAM ranks InceptionV3 above DenseNet201 is contradicted by the paper's own statistics. Table 2 reports InceptionV3 IoU 0.777 vs DenseNet201 0.699 (Δ=0.078, 11.2% relative), and §3.3 calls this an 'outperformance.' Appendix B Table 4 reports the same LayerCAM comparison as ΔIoU=-0.019, Cohen's d=-0.10, and Appendix B describes it as 'negligible effect' and 'minimal architectural differentiation.' These cannot both be true. Since the ranking reversal is the central evidence for the model-method-pair thesis, this inconsistency is fatal to the empirical demonstration as presented. The authors must supply the actual paired differences, confidence intervals, and one consistent set of aggregate tables, and then restate the reversal claim accordingly.
  2. [§3.3, Table 4, Table 5] The claimed effect sizes and cross-method deltas are inconsistent across the two statistical appendices. Section 3.3 states that InceptionV3's cross-method difference has 'Cohen's d = 1.47 (large effect),' but Appendix B Table 4 and Appendix C Table 5 both report d=0.73 for that comparison. Similarly, Table 4 gives DenseNet Layer-vs-Grad ΔIoU=0.015 (2.0%) and Inception Layer-vs-Grad ΔIoU=0.161 (21.2%), while Table 5 gives ΔIoU=0.009 (1.3%) and 0.134 (17.3%) for the same comparisons. The abstract, §3.3, §6.3, and Appendix B all quote different percentages. The authors need to reconcile these values and ensure that every claim cites the same underlying calculation.
  3. [Appendix G (Figures 3–5) vs Appendix E (Tables 7–8)] The qualitative visualization captions in Appendix G contradict the per-class quantitative tables. Figure 5 states that Grad-CAM++ for InceptionV3 causes COVID-19 IoU to fall 'from 0.949 to 0.071 (92.5% single-pathology collapse),' but Tables 7 and 8 report LayerCAM COVID-19 IoU=0.720 and Grad-CAM++ COVID-19 IoU=0.404 (a 43.9% reduction). Figure 5 also says Normal degrades '0.817→0.549,' whereas Tables 7 and 8 give 0.806→0.622. Figure 4 claims ResNet Normal 'catastrophically collapse[s] to 0.142,' while Table 7 reports ResNet LayerCAM Normal IoU=0.531 and Table 8 reports Grad-CAM++ Normal IoU=0.379. Either the captions refer to a different, unspecified processing pipeline (e.g., unthresholded or per-image IoU), or the tables are wrong. This must be clarified, because the qualitative panels are part of the evidence for the reversal claim.
  4. [§3.1 / Appendix A.1] The empirical study is not reproducible. The paper refers repeatedly to 'chest X-ray experiments' and Appendix A.1 gives class counts (3,354 test samples, 11,733 training, 1,675 validation), but the dataset is never named or linked, and no code is released. Since the paper's own controlled experiments are the professed empirical pillar of the position, and since the numerical inconsistencies above cannot be adjudicated without the underlying data, the manuscript should either provide the dataset identifier and analysis code or explicitly mark the experimental section as illustrative and reposition the argument entirely on the cited prior literature.
minor comments (5)
  1. [Abstract / §3.3] The phrase 'Grad-CAM++ favors DenseNet201 and reduced InceptionV3 stability score by 17.3%' is not consistent with Table 4's 21.2% and Table 5's 17.3% for the same comparison. Please harmonize the reported percentages.
  2. [Figure 2 caption vs §A.3] Figure 2's caption says the phase transition is from 'transfer learning (Epochs 1–9) to fine-tuning (Epochs 10–20),' but §A.3 defines Phase 1 as epochs 1–10 and Phase 2 as epochs 11–20. These should be aligned.
  3. [§7] The paper evaluates only two gradient-based attribution methods and a single dataset. Section 7 acknowledges this, but the abstract and conclusion present the reversal as a general finding. A sentence explicitly saying that the demonstration uses two gradient methods and one dataset, and that the broader thesis rests on prior literature, would better match the evidence.
  4. [§4.1 / Table 2] The text says LayerCAM implies 'DenseNet ≈ InceptionV3 > ResNet,' while §3.3 emphasizes InceptionV3's 11.2% advantage. These two framings are in tension. Once the numerical inconsistencies are resolved, the summary of the LayerCAM ordering should be stated consistently throughout.
  5. [Appendix B, Table 4] The table reports p<0.001 for DenseNet-vs-Inception LayerCAM with ΔIoU=-0.019 and d=-0.10. With N=2,430, a paired test can be significant with negligible effect size; this is fine statistically, but the accompanying text should not present it as supporting a meaningful ordering. The current wording ('small but significant advantage') is misleading.

Circularity Check

0 steps flagged

No significant circularity: the method-dependence claim is an empirical comparison, not a derivation from fitted parameters or self-cited assumptions.

full rationale

The paper's central claim—that explanation stability is a property of the model–method pair—rests on an empirical comparison of IoU-based semantic drift (Eq. 2) across two attribution methods (LayerCAM and Grad-CAM++) and three architectures. The stability score is defined directly as the overlap of thresholded attribution masks between transfer-learning and fine-tuned phases; this is a measurement choice, not a quantity fitted to predict the ranking. No parameter is fitted to the target ranking, and no 'prediction' is obtained from data already containing that ranking. The reversal conclusion (InceptionV3 highest under LayerCAM; DenseNet201 highest under Grad-CAM++) is presented as an observed result rather than derived from an assumption equivalent to it. The paper contains no self-citations by the present authors, and its supporting citations (Duan et al. 2024; Klein et al. LATEC; Hedström et al. Quantus; Arun et al.) are independent external evidence for method-dependence, not a self-citation chain. There is no imported uniqueness theorem and no ansatz smuggled via citation. The internal inconsistency between Table 2 (LayerCAM InceptionV3 IoU 0.777 vs DenseNet201 0.699) and Table 4 (DenseNet-vs-Inception ΔIoU = -0.019, d = -0.10) is a serious numerical/correctness problem that undermines confidence in the demonstration, but it is not circular reasoning. Likewise, the stated limitations in Section 7—only two gradient-based attribution methods, no dataset identifier or code, and metrics measuring consistency rather than clinical correctness—limit generalizability but do not make the argument circular. The position may be overgeneralized relative to its evidence, but the derivation chain is self-contained and not definitionally forced.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The central claims rest on measurement choices and filtering rules that are not independently validated: τ=0.2 threshold, epoch 8 vs 19 comparison, true-positive filtering, and the assumption that two gradient-based attribution methods represent method-dependence generally. The paper's own tables contain inconsistencies that make verification impossible without code/data.

free parameters (2)
  • IoU threshold τ = 0.2
    Hand-chosen binarization threshold for attribution masks; IoU values and rankings may change with τ, and no sensitivity analysis is provided (Section 3.1, Eq. 2).
  • Comparison epochs = 8 and 19
    Selected to maximize semantic drift contrast based on validation AUC plateau (Appendix A.4); results are specific to this phase pair.
axioms (5)
  • domain assumption LayerCAM and Grad-CAM++ are implemented correctly as specified in their original papers and applied to the penultimate layers.
    The empirical results inherit any implementation errors; no code is provided (Appendix A.5).
  • domain assumption IoU with τ=0.2 is a valid primary measure of explanation stability.
    Used as the primary metric (Eq. 2); other metrics (spatial displacement, pattern correlation, concentration) sometimes disagree, and no meta-rule selects IoU.
  • ad hoc to paper The true-positive-filtered sample (2,430 of 3,354) is representative for comparing architectures.
    The filter requires all three architectures to be correct in both phases (Eq. 6-8); samples where any model changed predictions are excluded, potentially selecting for unusually stable examples. No unfiltered sensitivity analysis is presented.
  • standard math The t-test on paired per-sample IoU differences is valid for these data.
    Assumes independent, roughly normal differences; with N=2430 significance is not informative about effect size, and the paper's own effect sizes are small for key comparisons (Appendix B).
  • domain assumption The no-free-lunch theorem for explanations (Han et al. 2022) applies and supports the normative claim.
    Cited in Sections 2.3 and 7 as theoretical grounding; the paper does not itself prove it.

pith-pipeline@v1.3.0-alltime-deepseek · 17598 in / 12945 out tokens · 111284 ms · 2026-08-01T20:18:48.790452+00:00 · methodology

0 comments
read the original abstract

This position paper argues that claims about explanation stability are scientifically invalid without cross method validation. Just as statistical significance requires the test statistic to be specified, stability should either be evaluated across multiple attribution paradigms or explicitly scoped to the computational objective of a single method. In controlled chest X ray experiments, DenseNet201, ResNet50V2, and InceptionV3 achieved AUC values above 99%, yet their stability rankings reversed across attribution methods. LayerCAM ranked InceptionV3 as the most stable model, with an IoU of 0.777, whereas GradCAM++ favored DenseNet201 and reduced InceptionV3 stability score by 17.3%. These findings demonstrate that explanation stability is an emergent property of the model method pair rather than an intrinsic characteristic of the model alone. We therefore argue that explanation based claims should be validated across multiple attribution methods and that regulatory submissions should explicitly specify the attribution operators used to avoid creating illusory safety assurances.

Figures

Figures reproduced from arXiv: 2607.16652 by Daniel Ting, Kabilan Elangovan.

Figure 1
Figure 1. Figure 1: Method-dependent stability rankings reveal complete architectural reversal. Left (LayerCAM): InceptionV3 (IoU=0.777) outperforms DenseNet201 (0.699) and ResNet50V2 (0.519). Right (Grad-CAM++): DenseNet201 (0.690) now leads InceptionV3 (0.643) and ResNet50V2 (0.383). InceptionV3 exhibits 17.3% method-dependent degradation; DenseNet maintains 1.3% cross-method consistency. Spatial displacement increases subs… view at source ↗
Figure 2
Figure 2. Figure 2: Epoch selection justification showing AUC progres￾sion across all 20 training epochs for three architectures. Epoch 8 (green circles) marks transfer learning plateau where valida￾tion AUC stabilizes at 0.969–0.996 across architectures, with DenseNet201 and ResNet50V2 showing minimal improvement thereafter (∆AUC<0.004 over next two epochs). Epoch 19 (red squares) represents fine-tuning convergence where all… view at source ↗
Figure 3
Figure 3. Figure 3: DenseNet201 cross-method comparison: LayerCAM (top) vs Grad-CAM++ (bottom). LayerCAM (IoU: 0.442–0.945): Exceptional stability with dense connectivity producing coherent explanation refinement. Normal fields achieve highest stability (0.945), COVID-19 maintains strong consistency (0.834), while Pneumonia represents the most dynamic adaptation (0.442). Grad-CAM++ (IoU: 0.467–0.946): Preserved cross-method s… view at source ↗
Figure 4
Figure 4. Figure 4: ResNet50V2 cross-method comparison: LayerCAM (top) vs Grad-CAM++ (bottom). LayerCAM (IoU: 0.393–0.630): Moderate instability with residual shortcuts enabling localized but unstable feature reconfiguration. Normal fields achieve highest relative stability (0.630), while Pneumonia collapses to lowest consistency (0.393). COVID-19 and Lung Opacity show mid-range performance (0.524, 0.595). Grad-CAM++ (IoU: 0.… view at source ↗
Figure 5
Figure 5. Figure 5: InceptionV3 cross-method comparison: LayerCAM (top) vs Grad-CAM++ (bottom). LayerCAM (IoU: 0.730–0.949): Excellent stability surpassing DenseNet in peak performance. COVID-19 achieves exceptional consistency (0.949); Lung Opacity main￾tains outstanding stability (0.905). Multi-scale pathway coherence produces uniformly high performance across all pathologies (min￾imum 0.730). Grad-CAM++ (IoU: 0.071–0.880):… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

58 extracted references · 5 linked inside Pith

  1. [1]

    Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages=

    Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization , author=. Proceedings of the IEEE International Conference on Computer Vision (ICCV) , pages=

  2. [2]

    2018 IEEE Winter Conference on Applications of Computer Vision (WACV) , pages=

    Grad-CAM++: Generalized Gradient-Based Visual Explanations for Deep Convolutional Networks , author=. 2018 IEEE Winter Conference on Applications of Computer Vision (WACV) , pages=. 2018 , organization=

    2018

  3. [3]

    IEEE Transactions on Image Processing , volume=

    LayerCAM: Exploring Hierarchical Class Activation Maps for Localization , author=. IEEE Transactions on Image Processing , volume=. 2021 , publisher=

    2021

  4. [4]

    International Conference on Machine Learning (ICML) , pages=

    Axiomatic Attribution for Deep Networks , author=. International Conference on Machine Learning (ICML) , pages=. 2017 , organization=

    2017

  5. [5]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    A Unified Approach to Interpreting Model Predictions , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  6. [6]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Sanity Checks for Saliency Maps , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  7. [7]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    A Benchmark for Interpretability Methods in Deep Neural Networks , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  8. [8]

    Explainable AI: Interpreting, Explaining and Visualizing Deep Learning , pages=

    The (Un)reliability of Saliency Methods , author=. Explainable AI: Interpreting, Explaining and Visualizing Deep Learning , pages=. 2019 , publisher=

    2019

  9. [9]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    On the (In)fidelity and Sensitivity of Explanations , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  10. [10]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Interpretation of Neural Networks is Fragile , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  11. [11]

    arXiv preprint arXiv:2011.08891 , year=

    Use HiResCAM instead of Grad-CAM for faithful explanations of convolutional neural networks , author=. arXiv preprint arXiv:2011.08891 , year=

    Pith/arXiv arXiv 2011

  12. [12]

    European Conference on Computer Vision (ECCV) , pages=

    Identity Mappings in Deep Residual Networks , author=. European Conference on Computer Vision (ECCV) , pages=. 2016 , organization=

    2016

  13. [13]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Densely Connected Convolutional Networks , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  14. [14]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Rethinking the Inception Architecture for Computer Vision , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  15. [15]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

    Network Dissection: Quantifying Interpretability of Deep Visual Representations , author=. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages=

  16. [16]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Interpreting Deep Visual Representations via Network Dissection , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2018 , publisher=

    2018

  17. [17]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Visualizing the Loss Landscape of Neural Nets , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  18. [18]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Transfusion: Understanding Transfer Learning for Medical Imaging , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  19. [19]

    Proceedings of the Conference on Health, Inference, and Learning (ACM CHIL) , pages=

    CheXtransfer: Performance and Parameter Efficiency of ImageNet Models for Chest X-Ray Interpretation , author=. Proceedings of the Conference on Health, Inference, and Learning (ACM CHIL) , pages=

  20. [20]

    Radiology: Artificial Intelligence , volume=

    RadImageNet: An Open Radiologic Deep Learning Research Dataset for Effective Transfer Learning , author=. Radiology: Artificial Intelligence , volume=. 2022 , publisher=

    2022

  21. [21]

    European Conference on Computer Vision (ECCV) , pages=

    Big Transfer (BiT): General Visual Representation Learning , author=. European Conference on Computer Vision (ECCV) , pages=. 2020 , organization=

    2020

  22. [22]

    International Conference on Machine Learning (ICML) , pages=

    Similarity of Neural Network Representations Revisited , author=. International Conference on Machine Learning (ICML) , pages=. 2019 , organization=

    2019

  23. [23]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Insights on Representational Similarity in Neural Networks with Canonical Correlation , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  24. [24]

    arXiv preprint arXiv:1711.05225 , year=

    CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning , author=. arXiv preprint arXiv:1711.05225 , year=

    Pith/arXiv arXiv

  25. [25]

    Medical Image Analysis , volume=

    Explainable Artificial Intelligence (XAI) in Deep Learning-Based Medical Image Analysis , author=. Medical Image Analysis , volume=. 2022 , publisher=

    2022

  26. [26]

    The Lancet Digital Health , volume=

    AI Recognition of Patient Race in Medical Imaging: A Modelling Study , author=. The Lancet Digital Health , volume=. 2022 , publisher=

    2022

  27. [27]

    Computer Methods and Programs in Biomedicine , volume=

    Explainable Deep Learning for Pulmonary Disease and Coronavirus COVID-19 Detection from X-rays , author=. Computer Methods and Programs in Biomedicine , volume=. 2020 , publisher=

    2020

  28. [28]

    Computers in Biology and Medicine , volume=

    On the evaluation of deep learning interpretability methods for medical images under the scope of faithfulness , author=. Computers in Biology and Medicine , volume=. 2024 , publisher=

    2024

  29. [29]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Do Vision Transformers See Like Convolutional Neural Networks? , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=

  30. [30]

    International Conference on Machine Learning (ICML) , pages=

    EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks , author=. International Conference on Machine Learning (ICML) , pages=. 2019 , organization=

    2019

  31. [31]

    Transactions on Machine Learning Research , year=

    The Disagreement Problem in Explainable Machine Learning: A Practitioner's Perspective , author=. Transactions on Machine Learning Research , year=

  32. [32]

    International Conference on Learning Representations (ICLR) , year=

    Towards Better Understanding of Gradient-Based Attribution Methods for Deep Neural Networks , author=. International Conference on Learning Representations (ICLR) , year=

  33. [33]

    Advances in Neural Information Processing Systems (NeurIPS) , volume=

    Which Explanation Should I Choose? A Function Approximation Perspective to Characterizing Post Hoc Explanations , author=. Advances in Neural Information Processing Systems (NeurIPS) , volume=. 2022 , note=

    2022

  34. [34]

    Journal of Machine Learning Research , volume=

    Quantus: An Explainable AI Toolkit for Responsible Evaluation of Neural Network Explanations and Beyond , author=. Journal of Machine Learning Research , volume=

  35. [35]

    The Meta-Evaluation Problem in Explainable

    Hedstr. The Meta-Evaluation Problem in Explainable. Transactions on Machine Learning Research , year =

  36. [36]

    2024 , eprint=

    On the Evaluation Consistency of Attribution-based Explanations , author=. 2024 , eprint=

    2024

  37. [37]

    ACM Computing Surveys , volume=

    From Anecdotal Evidence to Quantitative Evaluation Methods: A Systematic Review on Evaluating Explainable AI , author=. ACM Computing Surveys , volume=. 2023 , publisher=

    2023

  38. [38]

    Nature Machine Intelligence , volume=

    Benchmarking Saliency Methods for Chest X-ray Interpretation , author=. Nature Machine Intelligence , volume=. 2022 , publisher=

    2022

  39. [39]

    The Lancet Digital Health , volume=

    The False Hope of Current Approaches to Explainable Artificial Intelligence in Health Care , author=. The Lancet Digital Health , volume=. 2021 , publisher=

    2021

  40. [40]

    Science , volume=

    Beware Explanations from AI in Health Care , author=. Science , volume=. 2021 , publisher=

    2021

  41. [41]

    npj Digital Medicine , volume=

    Solving the Explainable AI Conundrum by Bridging Clinicians' Needs and Developers' Goals , author=. npj Digital Medicine , volume=. 2023 , publisher=

    2023

  42. [42]

    npj Digital Medicine , volume=

    The Clinician-AI Interface: Intended Use and Explainability in FDA-Cleared AI Devices for Medical Image Interpretation , author=. npj Digital Medicine , volume=. 2024 , publisher=

    2024

  43. [43]

    2021 , month=

    Good Machine Learning Practice for Medical Device Development: Guiding Principles , author=. 2021 , month=

    2021

  44. [44]

    2025 , month=

    Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence/Machine Learning (AI/ML)-Enabled Device Software Functions , author=. 2025 , month=

    2025

  45. [45]

    Radiology: Artificial Intelligence , volume=

    Assessing the Trustworthiness of Saliency Maps for Localizing Abnormalities in Medical Imaging , author=. Radiology: Artificial Intelligence , volume=. 2021 , publisher=

    2021

  46. [46]

    IEEE Transactions on Medical Imaging , volume=

    Attention Residual Learning for Skin Lesion Classification , author=. IEEE Transactions on Medical Imaging , volume=. 2019 , publisher=

    2019

  47. [47]

    ArXiv , year=

    OpenXAI: Towards a Transparent Evaluation of Model Explanations , author=. ArXiv , year=

  48. [48]

    arXiv preprint arXiv:2406.05090 , year=

    Provably Better Explanations with Optimized Aggregation of Feature Attributions , author=. arXiv preprint arXiv:2406.05090 , year=

    Pith/arXiv arXiv

  49. [49]

    Neurocomputing , volume=

    Drifting Explanations: On Explanation Stability in Continual Learning , author=. Neurocomputing , volume=. 2024 , publisher=

    2024

  50. [50]

    Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency , pages=

    Evaluation Gaps in Machine Learning Practice , author=. Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency , pages=

    2022

  51. [51]

    Patterns , volume=

    Leakage and the Reproducibility Crisis in ML-based Science , author=. Patterns , volume=. 2023 , publisher=

    2023

  52. [52]

    arXiv preprint arXiv:2502.06559 , year=

    Can We Trust AI Benchmarks? An Interdisciplinary Review of Current Issues in AI Evaluation , author=. arXiv preprint arXiv:2502.06559 , year=

    Pith/arXiv arXiv

  53. [53]

    arXiv preprint arXiv:1412.6980 , year=

    Adam: A Method for Stochastic Optimization , author=. arXiv preprint arXiv:1412.6980 , year=

    Pith/arXiv arXiv

  54. [54]

    2025 , eprint=

    Evaluating Explainability: A Framework for Systematic Assessment and Reporting of Explainable AI Features , author=. 2025 , eprint=

    2025

  55. [55]

    Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track , volume=

    Navigating the Maze of Explainable AI: A Systematic Approach to Evaluating Methods and Metrics , author=. Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track , volume=. 2024 , note=

    2024

  56. [56]

    2023 , publisher=

    Pirie, Craig and Wiratunga, Nirmalie and Wijekoon, Anjana , booktitle=. 2023 , publisher=

    2023

  57. [57]

    2025 , eprint=

    Benchmarking XAI Explanations with Human-Aligned Evaluations , author=. 2025 , eprint=

    2025

  58. [58]

    Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society , pages=

    Reckoning with the Disagreement Problem: Explanation Consensus as a Training Objective , author=. Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society , pages=

    2023