pith. sign in

arxiv: 2607.02394 · v1 · pith:DHSBXKVHnew · submitted 2026-07-02 · 📊 stat.AP · stat.ME

Masked complex non-decimated wavelet features for patient-level classification of contrast-enhanced mammography

Sara Antonijevic , Brani Vidakovic This is my paper

Pith reviewed 2026-07-03 02:32 UTC · model grok-4.3

classification 📊 stat.AP stat.ME
keywords contrast-enhanced spectral mammographywavelet featurespatient-level classificationAUC evaluationleakage-free cross-validationCESMelastic net
0
0 comments X

The pith

Low-energy and contrast CESM images reach statistically identical patient-level AUC of 0.874 under leakage-free evaluation, using disjoint wavelet channels.

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

The paper shows that a masked complex non-decimated wavelet feature set combined with elastic-net logistic regression produces patient-level malignancy detection on contrast-enhanced spectral mammography that matches the performance of a frozen ResNet-50 representation. It does so while keeping every coefficient directly tied to a physical image property such as phase coherence or magnitude distribution. The evaluation uses repeated patient-grouped nested cross-validation plus patient-cluster bootstrap inference on 1,880 images from 308 patients, removing the image-level leakage that has inflated earlier reported numbers. Under this protocol the low-energy and recombined contrast images turn out to be statistically indistinguishable, yet the malignancy signal travels through separate wavelet mechanisms in each acquisition. The result supplies an interpretable, reproducible baseline that future CESM classifiers can be compared against.

Core claim

Under single-image fusion the contrast image reaches a patient-level AUC of 0.874 (95% CI 0.827-0.918) and the low-energy image is statistically indistinguishable from it, yet the two encode malignancy through disjoint, interpretable channels: phase coherence on the low-energy image and magnitude distribution on the contrast image. The masked complex non-decimated wavelet representation with elastic-net logistic regression matches a pretrained ResNet-50 at the patient level while remaining fully interpretable at the level of individual predictors.

What carries the argument

Masked complex non-decimated wavelet feature bank that extracts phase-coherence and magnitude-distribution coefficients from each CESM image before elastic-net logistic classification.

If this is right

  • The low-energy and contrast acquisitions carry comparable overall signal strength but route that signal through different wavelet properties.
  • Every retained wavelet coefficient retains an explicit physical interpretation unavailable in the deep-network baseline.
  • The same evaluation protocol can serve as a standardized, leakage-free benchmark for any new CESM classifier.
  • Single-image fusion already suffices; multi-image fusion per patient is not required to reach the observed AUC level.

Where Pith is reading between the lines

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

  • Because the two channels are disjoint, a simple concatenation or weighted fusion of low-energy phase features with contrast magnitude features could raise AUC without adding model complexity.
  • The explicit physical meaning of each predictor makes it feasible to audit which spatial-frequency bands drive decisions on individual patients.
  • The same masked-wavelet pipeline could be tested on other dual-energy modalities such as dual-energy CT to check whether phase-versus-magnitude separation generalizes.

Load-bearing premise

Repeated patient-grouped nested cross-validation with patient-cluster bootstrap inference fully removes leakage and produces valid patient-level performance estimates on the CDD-CESM data.

What would settle it

Re-running the identical wavelet and ResNet models on the same 308-patient cohort but switching to image-level cross-validation folds and obtaining a materially higher AUC would indicate that the reported patient-level figures are still inflated by leakage.

Figures

Figures reproduced from arXiv: 2607.02394 by Brani Vidakovic, Sara Antonijevic.

Figure 1
Figure 1. Figure 1: Masked non-decimated two-dimensional wavelet decomposition of a mammogram. On [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Patient-cluster bootstrap 95% confidence intervals for image- and patient-level AUC across [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: ROC curves for the normal-versus-malignant task under the three modality configurations. [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Patient-cluster bootstrap 95% confidence intervals for image- and patient-level AUC [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Selection frequency of the top thirty stably selected features across outer-fold fits of [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Patient-level AUC under four fusion rules, for each of the three modality configurations. [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Patient-level ROC curves for the normal-versus-malignant task under max image fusion, [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Patient-level AUC under max image fusion for the masked complex NDWT feature bank (blue circles) and the ResNet-50 frozen-feature baseline (orange squares), on the three modality configurations. Bootstrap 95% CIs overlap substantially across all three configurations [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Wavelet vs ResNet-50 frozen features under matched patient-grouped cross-validation. [PITH_FULL_IMAGE:figures/full_fig_p022_9.png] view at source ↗
read the original abstract

Contrast-enhanced spectral mammography (CESM) acquires two images of each breast, a low-energy image and a recombined contrast image, but two questions central to building a classifier on them remain unsettled: whether the two image types carry comparable malignancy signal, and how a patient's several images should be combined into a single decision. Both are hard to answer reliably, because most published CESM classifiers split cross-validation folds at the image level, letting images of the same patient fall in both training and test sets and inflating reported performance. We pair a masked complex non-decimated wavelet feature bank with an elastic-net logistic classifier, evaluated under repeated patient-grouped nested cross-validation with patient-cluster bootstrap inference on the CDD-CESM dataset (1,880 images, 308 patients); under this leakage-free evaluation the inflation from testing on previously seen patients is negligible. On normal-versus-malignant detection, the two acquisitions are statistically indistinguishable in patient-level AUC under the proposed evaluation framework. Under single-image fusion the contrast image reaches a patient-level AUC of 0.874 (95% CI 0.827-0.918) and the low-energy image is statistically indistinguishable from it, yet the two encode malignancy through disjoint, interpretable channels: phase coherence on the low-energy image and magnitude distribution on the contrast image. The framework matches a pretrained ResNet-50 representation at the patient level, but whereas the frozen deep representation is not directly interpretable at the level of individual predictors, every predictor in the wavelet representation carries an explicit physical meaning. The result is a transparent, leakage-free baseline against which future CESM classifiers can be measured.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces masked complex non-decimated wavelet features combined with elastic-net logistic regression for patient-level normal-vs-malignant classification on contrast-enhanced spectral mammography (CESM). Using repeated patient-grouped nested cross-validation and patient-cluster bootstrap on the CDD-CESM dataset (1880 images, 308 patients), it reports that single-image fusion yields a patient-level AUC of 0.874 (95% CI 0.827-0.918) for the contrast image, statistically indistinguishable from the low-energy image; the two modalities rely on disjoint interpretable channels (phase coherence vs. magnitude distribution). The wavelet approach matches a frozen ResNet-50 at the patient level while remaining fully interpretable at the predictor level.

Significance. If the patient-grouped evaluation framework is correctly implemented, the work supplies a transparent, leakage-controlled baseline for CESM classification that demonstrates comparable signal strength in low-energy and contrast images through physically meaningful, disjoint wavelet features. This is valuable for clinical interpretability and for benchmarking future models against a non-deep-learning reference that avoids image-level leakage.

major comments (2)
  1. [Methods (cross-validation and bootstrap sections)] The validity of all reported AUCs, CIs, and the statistical equivalence between contrast and low-energy images rests on the patient-grouped nested CV plus cluster bootstrap. The manuscript must supply explicit pseudocode or a diagram (e.g., in §3 or §4) showing how every image belonging to a patient—including contralateral and multi-view acquisitions—is kept strictly out of the test fold, and how the bootstrap resamples at the patient level when re-estimating both elastic-net coefficients and AUC differences.
  2. [Methods (feature extraction and masking)] Wavelet scale selection and masking thresholds are listed among the tuned hyperparameters. The manuscript should clarify whether these choices are made inside the inner CV loop or fixed a priori; if the latter, any data-driven masking step performed on the full training set before outer-loop splitting risks subtle leakage that could affect the claimed disjoint-channel interpretation.
minor comments (2)
  1. [Methods] Notation for the complex wavelet coefficients and the masking operation should be introduced with a single consistent equation block rather than scattered across paragraphs.
  2. [Results] The exact number of bootstrap replicates used for the 95% CIs and the pairwise AUC comparison test should be stated explicitly.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight important aspects of reproducibility and leakage control in our patient-grouped evaluation framework. We address each major comment below and have revised the manuscript to incorporate the requested clarifications.

read point-by-point responses

  1. Referee: [Methods (cross-validation and bootstrap sections)] The validity of all reported AUCs, CIs, and the statistical equivalence between contrast and low-energy images rests on the patient-grouped nested CV plus cluster bootstrap. The manuscript must supply explicit pseudocode or a diagram (e.g., in §3 or §4) showing how every image belonging to a patient—including contralateral and multi-view acquisitions—is kept strictly out of the test fold, and how the bootstrap resamples at the patient level when re-estimating both elastic-net coefficients and AUC differences.

    Authors: We agree that explicit documentation is necessary to confirm the leakage-free nature of the evaluation. We will add pseudocode in Section 3 (and a supporting diagram in the supplement) that details the outer-loop patient grouping, the strict exclusion of all images (including contralateral and multi-view) from the same patient in test folds, the inner-loop hyperparameter tuning, and the patient-level cluster bootstrap procedure for re-estimating elastic-net coefficients and AUC differences. revision: yes

  2. Referee: [Methods (feature extraction and masking)] Wavelet scale selection and masking thresholds are listed among the tuned hyperparameters. The manuscript should clarify whether these choices are made inside the inner CV loop or fixed a priori; if the latter, any data-driven masking step performed on the full training set before outer-loop splitting risks subtle leakage that could affect the claimed disjoint-channel interpretation.

    Authors: All hyperparameter choices, including wavelet scale selection and masking thresholds, are performed strictly inside the inner cross-validation loop of the nested CV procedure. No data-driven steps occur on the full training set prior to outer-loop splitting. We will revise the Methods section to state this explicitly and add a sentence confirming that the inner loop isolates all tuning to prevent leakage, thereby preserving the validity of the disjoint-channel interpretation. revision: yes

Circularity Check

0 steps flagged

No circularity: patient-level AUCs computed on held-out patients via grouped CV, independent of fitted inputs.

full rationale

The paper's central results (AUC 0.874, statistical equivalence between image types) are obtained by applying an elastic-net classifier to masked complex non-decimated wavelet features and evaluating under repeated patient-grouped nested cross-validation with patient-cluster bootstrap on the 308-patient CDD-CESM dataset. No equations, ansatzes, or claims reduce these metrics by construction to the training data, fitted coefficients, or prior self-citations; the evaluation explicitly holds out entire patients. The derivation chain for feature construction and classification remains self-contained against external benchmarks, with no self-definitional loops, fitted-input predictions, or load-bearing self-citation chains.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 0 invented entities

The central claims rest on standard assumptions for clustered bootstrap inference and on the representativeness of the CDD-CESM dataset; hyperparameters for elastic net and wavelet scales are tuned but not enumerated as free parameters in the abstract.

free parameters (2)
  • elastic-net mixing parameter and penalty strength
    Tuned via nested CV; values not reported in abstract.
  • wavelet scale selection and masking thresholds
    Chosen to define the feature bank; specific values or selection procedure not given in abstract.
axioms (2)
  • domain assumption Patient grouping in CV fully prevents information leakage across images of the same patient
    Invoked when claiming the evaluation framework is leakage-free.
  • standard math Bootstrap inference on patient clusters produces valid 95% CIs for AUC
    Used to report confidence intervals around the 0.874 AUC.

pith-pipeline@v0.9.1-grok · 5830 in / 1498 out tokens · 34113 ms · 2026-07-03T02:32:45.569989+00:00 · methodology

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

54 extracted references · 1 canonical work pages

  1. [1]

    and Laversanne, Mathieu and Soerjomataram, Isabelle and Jemal, Ahmedin and Bray, Freddie , title =

    Sung, Hyuna and Ferlay, Jacques and Siegel, Rebecca L. and Laversanne, Mathieu and Soerjomataram, Isabelle and Jemal, Ahmedin and Bray, Freddie , title =. CA: A Cancer Journal for Clinicians , year =

  2. [2]

    and Giaquinto, Angela N

    Siegel, Rebecca L. and Giaquinto, Angela N. and Jemal, Ahmedin , title =. CA: A Cancer Journal for Clinicians , year =

  3. [3]

    and Guo, Helen and Martin, Lisa J

    Boyd, Norman F. and Guo, Helen and Martin, Lisa J. and Sun, Limei and Stone, Jennifer and Fishell, Eve and Jong, Roberta A. and Hislop, Greg and Chiarelli, Anna and Minkin, Salomon and Yaffe, Martin J. , title =. New England Journal of Medicine , year =

  4. [4]

    and Oestreicher, Nina and Porter, Peggy L

    Mandelson, Margaret T. and Oestreicher, Nina and Porter, Peggy L. and White, Donna and Finder, Charles A. and Taplin, Stephen H. and White, Emily , title =. Journal of the National Cancer Institute , year =

  5. [5]

    and Schmitzberger, Florian F

    Fallenberg, Eva M. and Schmitzberger, Florian F. and Amer, Hala and Ingold-Heppner, Barbara and Balleyguier, Corinne and Diekmann, Felix and Engelken, Florian and Mann, Ritse M. and Renz, Diane M. and Bick, Ulrich and Hamm, Bernd and Dromain, Clarisse , title =. European Radiology , year =

  6. [6]

    and Lobbes, Marc B

    Jochelson, Maxine S. and Lobbes, Marc B. I. , title =. Radiology , year =

  7. [7]

    and Lobbes, Marc B

    Patel, Bhavika K. and Lobbes, Marc B. I. and Lewin, John , title =. Seminars in Ultrasound, CT and MRI , year =

  8. [8]

    American Journal of Roentgenology , year =

    Sorin, Vera and Yagil, Yael and Yosepovich, Ady and Shalmon, Anat and Gotlieb, Michael and Neiman, Osnat Halshtok and Sklair-Levy, Miri , title =. American Journal of Roentgenology , year =

  9. [9]

    2021 , howpublished =

    Khaled, Rana and Helal, Maha and Alfarghaly, Omar and Mokhtar, Omnia and Elkorany, Abeer and El Kassas, Hebatalla and Fahmy, Aly , title =. 2021 , howpublished =. doi:10.7937/29kw-ae92 , url =

    work page doi:10.7937/29kw-ae92 2021

  10. [10]

    Scientific Data , year =

    Khaled, Rana and Helal, Maha and Alfarghaly, Omar and Mokhtar, Omnia and Elkorany, Abeer and El Kassas, Hebatalla and Fahmy, Aly , title =. Scientific Data , year =

  11. [11]

    Journal of Digital Imaging , year =

    Clark, Kenneth and Vendt, Bruce and Smith, Kirk and Freymann, John and Kirby, Justin and Koppel, Paul and Moore, Stephen and Phillips, Scott and Maffitt, Dennis and Pringle, Michael and Tarbox, Linda and Prior, Fred , title =. Journal of Digital Imaging , year =

  12. [12]

    Explainable Deep Learning for Breast Lesion Classification in Digital and Contrast-Enhanced Mammography , journal =

    Acosta-Jim. Explainable Deep Learning for Breast Lesion Classification in Digital and Contrast-Enhanced Mammography , journal =. 2025 , volume =

    2025

  13. [13]

    European Journal of Radiology , year =

    Helal, Maha and Khaled, Rana and Alfarghaly, Omar and Mokhtar, Omnia and Elkorany, Abeer and Fahmy, Aly and El Kassas, Hebatalla , title =. European Journal of Radiology , year =

  14. [14]

    Computerized Medical Imaging and Graphics , year =

    Gao, Fei and Wu, Teresa and Li, Jing and Zheng, Bin and Ruan, Lingxiang and Shang, Desheng and Patel, Bhavika , title =. Computerized Medical Imaging and Graphics , year =

  15. [15]

    Losurdo, Lorenzo and Fanizzi, Annarita and Basile, Teresa M. A. and Bellotti, Roberto and Bottigli, Ubaldo and Dentamaro, Rosalba and Didonna, Vittorio and Lorusso, Vito and Massafra, Raffaella and Tamborra, Pasquale and Tangaro, Sabina and La Forgia, Daniele , title =. Entropy , year =

  16. [16]

    and Pinker, Katja and Leithner, Doris and Sung, Janice and Avendano, Daly and Morris, Elizabeth A

    Marino, Maria A. and Pinker, Katja and Leithner, Doris and Sung, Janice and Avendano, Daly and Morris, Elizabeth A. and Jochelson, Maxine , title =. Molecular Imaging and Biology , year =

  17. [17]

    Statistical Modeling by Wavelets , year =

    Vidakovi. Statistical Modeling by Wavelets , year =

  18. [18]

    Ram. A 2. Computational Statistics & Data Analysis , year =

  19. [19]

    Mammogram Diagnostics via 2-

    Jeon, Seonghye and Nicolis, Orietta and Vidakovi. Mammogram Diagnostics via 2-. S. 2014 , volume =

    2014

  20. [20]

    Applied and Computational Harmonic Analysis , year =

    Kingsbury, Nick , title =. Applied and Computational Harmonic Analysis , year =

  21. [21]

    , title =

    Rouzrokh, Pouria and Khosravi, Bardia and Faghani, Shahriar and Moassefi, Mana and Vahdati, Sanaz and Erickson, Bradley J. , title =. Radiology: Artificial Intelligence , year =

  22. [22]

    Effect of data leakage in brain

    Yagis, Ekin and Atnafu, Selamawet Workalemahu and Garc. Effect of data leakage in brain. Scientific Reports , year =

  23. [23]

    Machine learning for medical imaging: methodological failures and recommendations for the future , journal =

    Varoquaux, Ga. Machine learning for medical imaging: methodological failures and recommendations for the future , journal =. 2022 , volume =

    2022

  24. [24]

    Patterns , year =

    Kapoor, Sayash and Narayanan, Arvind , title =. Patterns , year =

  25. [25]

    , title =

    Ambroise, Christophe and McLachlan, Geoffrey J. , title =. Proceedings of the National Academy of Sciences , year =

  26. [26]

    Journal of the Royal Statistical Society, Series B (Statistical Methodology) , year =

    Zou, Hui and Hastie, Trevor , title =. Journal of the Royal Statistical Society, Series B (Statistical Methodology) , year =

  27. [27]

    Journal of the Royal Statistical Society, Series B (Methodological) , year =

    Tibshirani, Robert , title =. Journal of the Royal Statistical Society, Series B (Methodological) , year =

  28. [28]

    and Kennard, Robert W

    Hoerl, Arthur E. and Kennard, Robert W. , title =. Technometrics , year =

  29. [29]

    and Olshen, Richard A

    Breiman, Leo and Friedman, Jerome H. and Olshen, Richard A. and Stone, Charles J. , title =. 1984 , publisher =

    1984

  30. [30]

    2009 , publisher =

    Hastie, Trevor and Tibshirani, Robert and Friedman, Jerome , title =. 2009 , publisher =

    2009

  31. [31]

    , title =

    Carneiro, Gustavo and Nascimento, Jacinto and Bradley, Andrew P. , title =. IEEE Transactions on Medical Imaging , year =

  32. [32]

    Deep neural networks improve radiologists' performance in breast cancer screening , journal =

    Wu, Nan and Phang, Jason and Park, Jungkyu and Shen, Yiqiu and Huang, Zhe and Zorin, Masha and Jastrz. Deep neural networks improve radiologists' performance in breast cancer screening , journal =. 2020 , volume =

    2020

  33. [33]

    and Ba, Jimmy Lei and Frey, Brendan J

    Kraus, Oren Z. and Ba, Jimmy Lei and Frey, Brendan J. , title =. Bioinformatics , year =

  34. [34]

    and Smithuis, Robin , title =

    Zonderland, Harmien M. and Smithuis, Robin , title =. 2014 , url =

    2014

  35. [35]

    and Reitsma, Johannes B

    Collins, Gary S. and Reitsma, Johannes B. and Altman, Douglas G. and Moons, Karel G. M. , title =. Annals of Internal Medicine , year =

  36. [36]

    The Annals of Statistics , year =

    Efron, Bradley , title =. The Annals of Statistics , year =

  37. [37]

    and Welsh, Alan H

    Field, Christopher A. and Welsh, Alan H. , title =. Journal of the Royal Statistical Society, Series B (Statistical Methodology) , year =

  38. [38]

    and DeLong, David M

    DeLong, Elizabeth R. and DeLong, David M. and Clarke-Pearson, Daniel L. , title =. Biometrics , year =

  39. [39]

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

    He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2016 , doi =

    2016

  40. [40]

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

    Deng, Jia and Dong, Wei and Socher, Richard and Li, Li-Jia and Li, Kai and Fei-Fei, Li , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages =. 2009 , doi =

    2009

  41. [41]

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

    Yosinski, Jason and Clune, Jeff and Bengio, Yoshua and Lipson, Hod , title =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  42. [42]

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

    Sharif Razavian, Ali and Azizpour, Hossein and Sullivan, Josephine and Carlsson, Stefan , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) , pages =. 2014 , doi =

    2014

  43. [43]

    Nature Machine Intelligence , year =

    Rudin, Cynthia , title =. Nature Machine Intelligence , year =

  44. [44]

    , title =

    Ghassemi, Marzyeh and Oakden-Rayner, Luke and Beam, Andrew L. , title =. The Lancet Digital Health , year =

  45. [45]

    Shortcut learning in deep neural networks , journal =

    Geirhos, Robert and Jacobsen, J. Shortcut learning in deep neural networks , journal =. 2020 , volume =

    2020

  46. [46]

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

    Roelofs, Rebecca and Shankar, Vaishaal and Recht, Benjamin and Fridovich-Keil, Sara and Hardt, Moritz and Miller, John and Schmidt, Ludwig , title =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  47. [47]

    and Kalpathy-Cramer, Jayashree , title =

    Arun, Nishanth and Gaw, Nathan and Singh, Praveer and Chang, Ken and Aggarwal, Mehak and Chen, Bryan and Hoebel, Katharina and Gupta, Sharut and Patel, Jay and Gidwani, Mishka and Adebayo, Julius and Li, Matthew D. and Kalpathy-Cramer, Jayashree , title =. Radiology: Artificial Intelligence , year =

  48. [48]

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

    Adebayo, Julius and Gilmer, Justin and Muelly, Michael and Goodfellow, Ian and Hardt, Moritz and Kim, Been , title =. Advances in Neural Information Processing Systems (NeurIPS) , volume =

  49. [49]

    Good Machine Learning Practice for Medical Device Development: Guiding Principles , year =

  50. [50]

    2024 , howpublished =

    Regulation (. 2024 , howpublished =

    2024

  51. [51]

    and Goldenberg, Anna , title =

    Tonekaboni, Sana and Joshi, Shalmali and McCradden, Melissa D. and Goldenberg, Anna , title =. Proceedings of the 4th Machine Learning for Healthcare Conference , series =

  52. [52]

    , title =

    Jha, Saurabh and Topol, Eric J. , title =. JAMA , year =

  53. [53]

    Winkler, Julia K. and Fink, Christine and Toberer, Ferdinand and Enk, Alexander and Deinlein, Teresa and Hofmann-Wellenhof, Rainer and Thomas, Luc and Lallas, Aimilios and Blum, Andreas and Stolz, Wilhelm and Haenssle, Holger A. , title =. JAMA Dermatology , year =

  54. [54]

    and Janizek, Joseph D

    DeGrave, Alex J. and Janizek, Joseph D. and Lee, Su-In , title =. Nature Machine Intelligence , year =