Pith. sign in

REVIEW 4 major objections 5 minor 28 references

A Diagnostic Gap Framework for Evaluating Reconstruction Fidelity in Weakly Supervised Mammography

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Diagnostic loss in reconstruction tracks fidelity, not architecture

desk verdict A genuinely useful evaluation framework with a solid fidelity ladder; the frozen-probe OOD concern is real but the dose-response sweep largely disarms it, and the paper needs minor corrections before acceptance. read the letter →

arxiv 2607.22740 v1 pith:2P3GQGRL submitted 2026-07-22 cs.CV

Vinceline Bertrand , Ionut Cardei This is my paper
classification cs.CV
keywords diagnosticgapreconstructionfidelityweaklysupervisedmammographydecisionpreservationexplanationattributionsimilarityLPIPSSDEdit
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

This paper tries to establish that the loss of diagnostic information when mammograms pass through generative reconstruction is not an inevitable cost of reconstruction but a controllable function of how faithful the reconstruction is. Using a frozen classifier as a fixed measurement instrument, the authors measure both decision preservation (AUC, accuracy, probability shifts) and explanation preservation (attribution map similarity) across a fidelity ladder of three reconstructors spanning a twenty-fold range of perceptual distance. At autoencoder-level fidelity, diagnostics are preserved nearly perfectly; at diffusion-level fidelity, both decisions and explanations collapse, with mass-pathology AUC falling below chance. The paper argues that the 'diagnostic gap' can therefore be read off a fidelity axis rather than attributed to the choice of reconstruction architecture.

What carries the argument

A frozen diagnostic probe — a calibrated classifier fixed after training and applied unchanged to original and reconstructed ROI crops — serves as the measurement instrument, ensuring any output change is attributable to reconstruction rather than retraining. A fidelity ladder of three class-conditional reconstructors (discrete-latent VQ-VAE-GAN, continuous-latent VAE-GAN, and diffusion via SDEdit) spans LPIPS 0.029–0.584 and provides the independent fidelity axis, while dual decision and explanation preservation metrics (agreement, ΔAcc, ΔAUC, |Δp|, and HiResCAM/Grad-CAM++ cosine/SSIM) capture both whether predictions change and whether the evidence shifts. The task-level diagnostic gap Δdi

What would settle it

Retrain the diagnostic probe on reconstructed images (or fine-tune it with reconstructions) and check whether AUC on low-fidelity reconstructions recovers: if it does, the original diagnostic gap overstates information loss. Additionally, apply attribution sanity checks (e.g., randomizing the probe's weights and observing whether saliency maps change on reconstructed inputs) to test whether attribution similarity scores genuinely reflect preserved evidence.

Watch

Extended reading notes

Core claim

The diagnostic gap is a measurable function of reconstruction fidelity rather than an intrinsic cost of reconstruction. At LPIPS ≤ 0.078 (autoencoder fidelity), the frozen probe's AUC changes by at most ±0.005 and attribution similarity remains high; at LPIPS 0.584 (diffusion SDEdit fidelity), pooled pathology AUC drops by 0.253 and mass-pathology AUC falls to 0.461, below chance. A seven-level SDEdit strength sweep confirms monotonic degradation: as perceptual distance increases, AUC decreases steadily from near zero to −0.292, establishing a dose-response relationship between reconstruction fidelity and diagnostic preservation.

Load-bearing premise

The frozen classifier trained on original lesion images remains a trustworthy measurement instrument on reconstructed images; if reconstructions make the classifier's decisions or saliency maps unreliable for reasons other than actually destroying clinical information, the measured diagnostic gap would partly reflect probe out-of-distribution behavior rather than pure information loss.

Editorial extensions

If this is right

  • If the central claim holds, multi-stage medical AI pipelines can be evaluated stage-wise: a high-fidelity reconstructor incurs little diagnostic cost, so reconstruction is not inherently dangerous.
  • The framework produces a reusable dose-response curve, allowing new reconstructors to be placed on the same fidelity axis and compared without treating each architecture as a separate evaluation setting.
  • Explanation preservation is more sensitive than decision preservation: a pipeline can preserve the correct label while shifting the classifier's attention, which matters in regulatory settings where explanations must accompany predictions.
  • Coarse diagnostic tasks (e.g., lesion type) are more robust to reconstruction degradation than fine-grained tasks (e.g., pathology), so high coarse-task accuracy can overstate the reliability of a pipeline's fine-grained diagnostic information.
  • The monotonic SDEdit strength sweep implies that diagnostic degradation is not binary but graded, and even minimal diffusion strength incurs a baseline fidelity cost that may already exceed the autoencoder range.

Reading between the lines

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

  • A direct testable extension: if the fidelity-dependence claim is right, then training a larger or pretrained diffusion model to reach autoencoder-level LPIPS should recover diagnostic preservation; if it still degrades, the gap would be partly architecture-intrinsic rather than purely fidelity-driven.
  • The decision-explanation dissociation suggests that evaluation protocols for medical AI should require both decision and explanation preservation metrics, because label preservation alone can hide shifts in the evidence base the classifier uses.
  • The framework naturally extends upstream to localization error, suggesting a compound pipeline: localization plus reconstruction may degrade diagnostics faster than either alone, and the diagnostic gap metric could quantify that interaction.
  • Explanation preservation could be incorporated as a training objective for reconstructors, steering them to preserve diagnostic attention rather than only pixel-level or perceptual similarity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a 'diagnostic gap' evaluation framework for weakly supervised mammography pipelines. Using a frozen ResNet-50 classifier as a fixed probe, the authors measure decision preservation (agreement, ΔAcc, ΔAUC, |Δp|) and explanation preservation (HiResCAM/Grad-CAM++ similarity) on lesion ROI crops reconstructed by VQ-VAE-GAN, VAE-GAN, and diffusion (SDEdit) at varying strengths. They report that at autoencoder fidelity (LPIPS ≤ 0.078) the diagnosis is preserved (AUC within ±0.005), while at diffusion fidelity (LPIPS 0.584) AUC drops sharply, with mass pathology below chance, and conclude that the diagnostic gap is a function of reconstruction fidelity rather than an intrinsic cost of reconstruction.

Significance. If the central claim is established, the framework is a practical, architecture-agnostic tool for stage-wise evaluation of medical image pipelines. The paper has notable strengths: a controlled fidelity ladder, a seven-point SDEdit strength sweep, bootstrap confidence intervals, and complementary decision/explanation metrics anchored by a frozen probe. The dose-response relationship between LPIPS and ΔAUC is a useful and falsifiable observation. However, the central interpretation depends critically on the assumption that the frozen probe remains a valid diagnostic instrument on reconstructed inputs; this is not established, and two quantitative claims in the abstract are not supported by the paper's own tables.

major comments (4)
  1. [Abstract; Table I] The abstract and contributions state that at autoencoder fidelity 'AUC changes remain within ±0.005.' Table I shows the lesion-type VQ-VAE-GAN row: ΔAUC = -0.007 with 95% CI [-0.012, -0.002], which is outside ±0.005 and excludes zero. Please either restrict the claim to pathology tasks or report the actual range (e.g., within ±0.007). This is not merely cosmetic because the CI excludes zero, contradicting the statement that the effect is within measurement noise.
  2. [Table I vs Table III] The primary evaluation uses SDEdit at strength 0.5 (Sec IV-D). Table I reports pooled pathology ΔAUC = -0.253 for this operating point, while Table III reports ΔAUC = -0.227 for the same strength and LPIPS (0.584). The abstract and Sec V.C repeat the -0.253 value. Please reconcile this discrepancy; if the difference arises from bootstrap sampling, state the resampling procedure and report the CI in both places.
  3. [§IV-C, §III-F, §VI-B] The central claim that the diagnostic gap is a function of fidelity rather than of the probe's out-of-distribution brittleness assumes that the frozen ResNet-50 probe remains a valid instrument on reconstructions. Calibration and conformal coverage are validated only on original validation ROIs (§IV-C), and attribution maps on reconstructed inputs are not sanity-checked (§III-F). Under SDEdit, the input distribution shifts substantially (LPIPS 0.584, SSIM 0.408), so the observed ΔAUC = -0.253 and mass-pathology below-chance AUC could reflect loss of features the probe depends on, rather than loss of clinically relevant information. Please provide a feature-recovery or probe-reliability check, e.g., train a probe on reconstructed ROIs and measure transfer, or verify that reconstruction preserves the statistics needed for a held-out reader/probe. Without this, the Discussion VI-B interpret
  4. [Table I, mass pathology] The statement that mass-pathology AUC 'falls below chance' is based on a point estimate of 0.461. The bootstrap CI for ΔAUC is [-0.422, -0.261]; combined with the original AUC of 0.802, the reconstructed AUC has CI [0.380, 0.541], which includes 0.5. Thus the data do not support a statistically significant below-chance result. Please qualify this claim in the abstract and results, or compute the CI directly for the reconstructed AUC.
minor comments (5)
  1. [§VI.C] The sentence beginning 'The diagnostic gap metric Δdiag...' appears to be a fragment after a duplicated line; the preceding sentence is cut off.
  2. [§IV.D] The statement 'A separate reconstructor is trained per cascade stage' is not described—what does 'per cascade stage' mean? Do the autoencoder and diffusion reconstructors differ per stage, and are the same reconstructions used for both the lesion-type and pathology probes?
  3. [Figure 5] The legend lists strength values 0.05, 0.2, 0.5, 0.7 only for four curves, but the text describes seven levels; the curves are not fully identified. Consider plotting all seven or labeling explicitly.
  4. [References] References [19] and [7] are dated 2026; please confirm these are published/available.
  5. [Throughout] Typos: 'fideliy' in §III; inconsistent spacing in 'V AE-GAN' in several places. Table II lacks bootstrap CIs for attribution similarity; adding them would strengthen the claim that the ranking is reliable.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: fidelity and preservation are independently measured; the claim is empirical and self-contained.

full rationale

The paper's central relationship is between measured reconstruction fidelity (LPIPS, SSIM) and diagnostic preservation (ΔAUC, agreement, |Δp|, attribution similarity). These are not definitionally linked: LPIPS/SSIM compare original and reconstructed pixel/perceptual structure, while ΔAUC and attribution similarity are computed by a frozen ResNet-50 probe on original versus reconstructed ROIs. No parameter is fitted from one axis to predict the other, and the SDEdit strength sweep (0.05–0.7) provides an external manipulation that traces ΔAUC monotonically against LPIPS within a single reconstructor, so the fidelity-dependence conclusion is not forced by construction. All cited methods (SDEdit, HiResCAM, Grad-CAM++, temperature scaling, conformal prediction, CBIS-DDSM) are external; there are no self-citations or imported uniqueness theorems that carry the load. The manuscript's acknowledged limitations (§VI-D: ROI-only evaluation, diffusion model capacity, dataset/evaluator scope, attribution methods) bound the generality of the result but do not introduce circularity. The frozen-probe transfer to reconstructions is a measurement-validity assumption that could affect correctness if reconstructed inputs are out-of-distribution for the probe, but that is an empirical risk, not a definitional reduction of the dependent variable to the independent one.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim is an empirical correlation, so the ledger is dominated by measurement assumptions rather than fitted constants. The main hand-chosen operating points are SDEdit strength 0.5 and the training budgets that place each reconstructor on the fidelity axis. No invented physical entities; the 'diagnostic gap' is a computed metric (Eq. 1).

free parameters (3)
  • SDEdit strength (primary diffusion operating point) = 0.5
    Chosen for the headline diffusion comparison; at strength 0.05 ΔAUC≈0, so the reported collapse depends on this operating point.
  • Reconstructor training budget/loss weights = 10 or 30 epochs; 100·L_MSE + L_commit + 2·L_perc; adversarial weight 0→0.5
    Hand-selected; determines each reconstructor's LPIPS and thus its place on the fidelity axis. No sensitivity analysis reported.
  • Probe class weights and training schedule = inverse-frequency weights; AdamW LR 1e-4, wd 1e-4, 10-25 epochs
    Chosen to create a usable probe; probe sensitivity sets the scale of ΔAUC and |Δp|, but it is frozen across all rungs.
assumptions (5)
  • domain assumption LPIPS and SSIM adequately capture clinically relevant reconstruction fidelity
    The fidelity axis and the 'function of fidelity' claim use LPIPS/SSIM as the independent variable (§III-C); if they miss clinically relevant morphology, the dose-response curve is miscalibrated.
  • domain assumption A frozen classifier trained on original ROIs remains valid on reconstructed inputs
    Calibration and conformal validation are done only on original validation ROIs (§IV-C), yet the probe's ΔAUC, |Δp|, and attribution maps on reconstructions are read as true diagnostic-information loss.
  • domain assumption Gradient-based attribution maps on reconstructed images reflect the probe's evidence
    Explanation preservation is measured with HiResCAM/Grad-CAM++ similarity (§III-F, §V-D), but no sanity checks or baselines are run on reconstructions; attribution reliability at low fidelity is assumed.
  • domain assumption CBIS-DDSM curated ROI crops are a sufficient testbed
    Single dataset, 3,550 ROIs total, 520 test, one scanner family; generalization to other populations/modalities is untested and acknowledged in §VI-D.
  • domain assumption SDEdit strength sweep with one diffusion model isolates fidelity from architecture
    The causal reading ('fidelity, not intrinsic') leans on the within-model strength sweep (§V-F); the autoencoder-vs-diffusion comparison also changes capacity, training, and latent structure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Diagnostic Gap Framework for Evaluating Reconstruction Fidelity in Weakly Supervised Mammography." pith.science (2026). https://pith.science/paper/2P3GQGRL

@misc{pith2026260722740,
  author       = {Pith},
  title        = {Pith review of: A Diagnostic Gap Framework for Evaluating Reconstruction Fidelity in Weakly Supervised Mammography},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2P3GQGRL}},
  note         = {Machine review of arXiv:2607.22740}
}
abstract

Weakly supervised pipelines for medical imaging have become increasingly popular over the years. These systems often include multiple stages and components, such as reconstruction, generation, and localization, yet standard evaluation metrics provide limited insight into whether clinically relevant information is preserved across each stage. We present the diagnostic gap framework, a practical evaluation tool that measures decision preservation and explanation preservation as a function of measured reconstruction fidelity. To isolate the effect of reconstruction from localization, we evaluate on curated lesion ROI crops using a fidelity ladder of three class-conditional reconstructors---VQ-VAE-GAN, VAE-GAN, and diffusion (SDEdit)---spanning a twenty-fold range in perceptual distance (LPIPS 0.029--0.584). At autoencoder fidelity, both decision and explanation are preserved: AUC changes remain within $\pm$0.005 and attribution similarity (HiResCAM, Grad-CAM++) stays high. At diffusion fidelity, both collapse: pooled AUC drops by 0.253 and mass-pathology AUC falls below chance. The diagnostic gap is thus a measurable function of reconstruction fidelity rather than an intrinsic cost of reconstruction, and the framework provides an architecture-agnostic instrument for identifying when and where multi-stage pipelines lose diagnostic signal.

Figures

Figures reproduced from arXiv: 2607.22740 by the authors.

Figure 2
Figure 2. Empirical coverage as a function of target coverage [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Decision preservation (∆AUC) as a function of reconstruction fidelity (LPIPS). At autoencoder fidelity (low LPIPS), AUC is preserved within ±0.005. At diffusion fidelity (high LPIPS), AUC degrades sharply, with mass pathology falling below chance. signed shift is positive, indicating SDEdit biases predictions toward malignant rather than scattering symmetrically. Prediction agreement is the least informative measure… view at source ↗
Figure 4
Figure 4. Attribution preservation across the fidelity ladder. Both [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Dose-response curve: ∆AUC as a function of LPIPS across SDEdit strengths (0.05–0.7) and autoencoder baselines. Degradation increases monotonically with perceptual distance across all subsets. Mass pathology is most vulnerable, dropping below chance at strength ≥0.5. fi…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 2 linked inside Pith

  1. [1]

    Learning deep features for discriminative localization,

    B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Learning deep features for discriminative localization,”Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016

    2016

  2. [2]

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

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-cam: Visual explanations from deep networks via gradient-based localization,” inProceedings of the IEEE International Conference on Computer Vision (ICCV), 2017

    2017

  3. [3]

    Neural discrete representation learning,

    A. van den Oord, O. Vinyals, and K. Kavukcuoglu, “Neural discrete representation learning,” inAdvances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  4. [4]

    Diffusion models for medical image analysis: A comprehensive survey,

    A. Kazerouni, E. Khodapanah Aghdam, M. Heidari, R. Azad, M. Fayyaz, I. Hacihaliloglu, and D. Merhof, “Diffusion models for medical image analysis: A comprehensive survey,”Medical Image Analysis, 2023. Covers diffusion-based methods across medical imaging applications

    2023

  5. [5]

    Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs,

    J. R. Zech, M. A. Badgeley, M. Liu, A. B. Costa, J. J. Titano, and E. K. Oermann, “Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs,”PLoS Medicine, vol. 15, no. 11, p. e1002683, 2018

    2018

  6. [6]

    Breast cancer survival rates

    A. C. Society, “Breast cancer survival rates.” https://www.cancer.org/ cancer/types/breast-cancer/understanding-a-breast-cancer-diagnosis/ breast-cancer-survival-rates.html, 2025. Accessed: 2026-01

    2025

  7. [7]

    Why we still miss breast cancers: Strategies for improving mammography interpretation,

    N. Chotai, A. Gadwal, D. Buchireddy, and W. T. Yang, “Why we still miss breast cancers: Strategies for improving mammography interpretation,” Insights into Imaging, vol. 17, no. 1, p. 8, 2026

    2026

  8. [8]

    Taming transformers for high- resolution image synthesis,

    P. Esser, R. Rombach, and B. Ommer, “Taming transformers for high- resolution image synthesis,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021

    2021

Show all 28 references
  1. [9]

    Autoencoding beyond pixels using a learned similarity metric,

    A. B. L. Larsen, S. K. Sønderby, H. Larochelle, and O. Winther, “Autoencoding beyond pixels using a learned similarity metric,” in International Conference on Machine Learning (ICML), 2016

    2016

  2. [10]

    SDEdit: Guided image synthesis and editing with stochastic differential equations,

    C. Meng, Y . He, Y . Song, J. Song, J. Wu, J.-Y . Zhu, and S. Ermon, “SDEdit: Guided image synthesis and editing with stochastic differential equations,” inInternational Conference on Learning Representations (ICLR), 2022

    2022

  3. [11]

    Deep learning to improve breast cancer detection on screening mammography,

    L. Shen, L. R. Margolies, J. H. Rothstein, E. Fluder, R. McBride, and W. Sieh, “Deep learning to improve breast cancer detection on screening mammography,”Scientific Reports, vol. 9, no. 1, p. 12495, 2019

    2019

  4. [12]

    Breast cancer detection in mammography using a convolu- tional neural network,

    D. Yi, S. Sawyer, A. Raghunathan, C. Chennubhotla, and S. Krish- naswamy, “Breast cancer detection in mammography using a convolu- tional neural network,”Medical Physics, vol. 44, no. 8, pp. 4287–4297, 2017

    2017

  5. [13]

    Patient-specific mri super-resolution via implicit neural representations,

    Y . Liet al., “Patient-specific mri super-resolution via implicit neural representations,”Medical Image Analysis, 2025

    2025

  6. [14]

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

    R. L. Draelos and L. Carin, “Use HiResCAM instead of Grad-CAM for faithful explanations of convolutional neural networks,” inarXiv preprint arXiv:2011.08891, 2021

    2011 arXiv

  7. [15]

    Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks,

    A. Chattopadhay, A. Sarkar, P. Howlader, and V . N. Balasubramanian, “Grad-CAM++: Generalized gradient-based visual explanations for deep convolutional networks,” inIEEE Winter Conference on Applications of Computer Vision (WACV), pp. 839–847, 2018

    2018

  8. [16]

    On calibration of modern neural networks,

    C. Guo, G. Pleiss, Y . Sun, and K. Q. Weinberger, “On calibration of modern neural networks,”International Conference on Machine Learning (ICML), 2017

    2017

  9. [17]

    Beyond temperature scaling: Ob- taining well-calibrated multiclass probabilities with dirichlet calibration,

    M. Kull, T. Silva Filho, and P. Flach, “Beyond temperature scaling: Ob- taining well-calibrated multiclass probabilities with dirichlet calibration,” Advances in Neural Information Processing Systems, 2019

    2019

  10. [18]

    A gentle introduction to conformal prediction and distribution-free uncertainty quantification,

    A. N. Angelopoulos and S. Bates, “A gentle introduction to conformal prediction and distribution-free uncertainty quantification,”arXiv preprint arXiv:2107.07511, 2021

    2021 arXiv

  11. [19]

    Cmt-unet: Leveraging stage-wise hybrid framework for enhanced accuracy and efficiency in medical image segmentation,

    R. Wang, H. Liu, and G. Wang, “Cmt-unet: Leveraging stage-wise hybrid framework for enhanced accuracy and efficiency in medical image segmentation,”Scientific Reports, vol. 16, p. 10079, 2026

    2026

  12. [20]

    Sanity checks for saliency maps,

    J. Adebayo, J. Gilmer, M. Muelly, I. Goodfellow, M. Hardt, and B. Kim, “Sanity checks for saliency maps,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 31, 2018

    2018

  13. [21]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016

    2016

  14. [22]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 586–595, 2018

    2018

  15. [23]

    A curated mammography data set for use in computer-aided detection and diagnosis research,

    R. S. Lee, F. Gimenez, A. Hoogi, and D. Rubin, “A curated mammography data set for use in computer-aided detection and diagnosis research,” Scientific Data, vol. 4, p. 170177, 2017

    2017

  16. [24]

    Inbreast: Toward a full-field digital mammographic database,

    I. C. Moreira, I. Amaral, I. Domingues, A. Cardoso, M. J. Cardoso, and J. S. Cardoso, “Inbreast: Toward a full-field digital mammographic database,”Academic Radiology, vol. 19, no. 2, pp. 236–248, 2012

    2012

  17. [25]

    Vindr-mammo: A large-scale benchmark dataset for computer- aided diagnosis in full-field digital mammography,

    H. T. Nguyen, H. Q. Nguyen, H. H. Pham, K. Lam, L. T. Le, M. Dao, and V . Vu, “Vindr-mammo: A large-scale benchmark dataset for computer- aided diagnosis in full-field digital mammography,”Scientific Data, vol. 10, no. 1, p. 277, 2023

    2023

  18. [26]

    RISE: Randomized input sampling for explanation of black-box models,

    V . Petsiuk, A. Das, and K. Saenko, “RISE: Randomized input sampling for explanation of black-box models,” inProceedings of the British Machine Vision Conference (BMVC), 2018

    2018

  19. [27]

    Visualizing and understanding convolutional networks,

    M. D. Zeiler and R. Fergus, “Visualizing and understanding convolutional networks,” inECCV, pp. 818–833, Springer, 2014

    2014

  20. [28]

    Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (TCA V),

    B. Kim, M. Wattenberg, J. Gilmer, C. Cai, J. Wexler, F. Viegas, and R. Sayres, “Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (TCA V),” inICML, pp. 2668–2677, 2018

    2018

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.