Pith. sign in

REVIEW 3 major objections 5 minor 65 references

A multi-center breast MRI benchmark finds tumor segmentation generalizes across continents, while pre-treatment prediction of pathologic complete response barely exceeds random.

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

2026-08-02 19:39 UTC pith:ZSV56MII

load-bearing objection A well-run challenge paper with a genuinely important negative pCR finding, but the blanket conclusion needs per-center analysis and the fairness metrics need uncertainty quantification. the 3 major comments →

arxiv 2603.01250 v3 pith:ZSV56MII submitted 2026-03-01 cs.CV cs.AI

The MAMA-MIA Challenge: Advancing Generalizability and Fairness in Breast MRI Tumor Segmentation and Treatment Response Prediction

Lidia Garrucho , Smriti Joshi , Kaisar Kushibar , Richard Osuala , Maciej Bobowicz , Xavier Bargall\'o , Paulius Jaru\v{s}evi\v{c}ius , Kai Geissler
show 38 more authors
Raphael Sch\"afer Muhammad Alberb Tony Xu Anne Martel Daniel Sleiman Navchetan Awasthi Hadeel Awwad Joan C. Vilanova Robert Mart\'i Daan Schouten Jeong Hoon Lee Mirabela Rusu Eleonora Poeta Luisa Vargas Eliana Pastor Maria A. Zuluaga Jessica K\"achele Dimitrios Bounias Alexandra Ertl Katarzyna Gwo\'zdziewicz Maria-Laura Cosaka Pasant M. Abo-Elhoda Sara W. Tantawy Shorouq S. Sakrana Norhan O. Shawky-Abdelfatah Amr Muhammad Abdo-Salem Androniki Kozana Eugen Divjak Gordana Ivanac Katerina Nikiforaki Michail E. Klontzas Rosa Garc\'ia-Dosd\'a Meltem Gulsun-Akpinar O\u{g}uz Lafc{\i} Carlos Mart\'in-Isla Oliver D\'iaz Laura Igual Karim Lekadir
This is my paper
classification cs.CV cs.AI
keywords breast MRIDCE-MRItumor segmentationpathologic complete responsegeneralizabilityalgorithmic fairnessbenchmark challengeneoadjuvant chemotherapy
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.

The paper reports results of a large benchmark designed to test whether AI models trained on US breast DCE-MRI can generalize to European centers and perform consistently across demographic subgroups. For tumor segmentation, the answer is largely yes: top submissions improved on a strong baseline in both accuracy and fairness, with main failures confined to small, low-contrast, non-mass lesions. For prediction of pathologic complete response from pre-treatment imaging alone, the answer is largely no: the best submitted models scored barely above a random classifier, calibration was poor, and only one submission was statistically distinguishable from chance. The authors conclude that pre-treatment pCR prediction from a single baseline DCE-MRI is not clinically reliable in a multi-center setting and that previously published single-center successes should be viewed with caution. They also show that the balance between accuracy and subgroup fairness is a policy choice: changing the fairness weight reshuffles leaderboard rankings substantially.

Core claim

The paper's central claim is that under external, cross-continental evaluation, automated breast tumor segmentation is robust enough to consider for multi-center deployment, while pCR prediction from baseline DCE-MRI is not. On a held-out test set of 574 patients from three European centers, top segmentation submissions improved over a strong baseline by up to roughly five Dice points while also improving subgroup fairness. In contrast, the best pCR submissions achieved balanced accuracies near 0.50, statistically indistinguishable from a random classifier for all but one of fourteen teams; calibration was poor and recall for the pCR class was low. Because accurate segmentations did not tran

What carries the argument

The load-bearing instrument is the benchmark's evaluation design: a single combined score S = (1-λ)Sp + λSf with λ=0.5, where Sp is task accuracy (Dice and normalized Hausdorff distance for segmentation; balanced accuracy for pCR) and Sf measures consistency across subgroups defined by age, menopausal status, and breast density. Paired with a private external test set of 574 fat-suppressed DCE-MRI exams from three European centers and a US multi-institution training cohort of 1,506 patients, this design forces models trained on US data to be tested on unseen scanners and protocols, and re-ranks submissions once fairness is included. The random-classifier baseline for pCR sets a lower bound t

Load-bearing premise

The load-bearing premise is that the 574-patient external test cohort from three European centers is representative enough of the broader clinical population that the near-random pCR result reflects a property of pre-treatment DCE-MRI itself, rather than an artifact of this particular sample's tumor sizes, pCR prevalence, or scanner protocols.

What would settle it

Run the same public training set against a different external cohort with more balanced pCR prevalence (or enriched for subtypes with higher pCR rates, such as triple-negative or HER2-positive tumors) and check whether a model achieves balanced accuracy clearly above 0.60 with well-calibrated probabilities. If such a model appears while this benchmark's test set stays near random, the paper's 'no reliable pre-treatment signal' conclusion would be weakened to 'no reliable signal under this test distribution.'

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Tumor segmentation models trained on diverse multi-institutional data can generalize to new continents and scanners; top methods beat a strong baseline on both accuracy and fairness.
  • Fairness-aware evaluation changes conclusions: performance gaps across subgroups are modest for segmentation, and rankings shift materially as the accuracy–fairness weight λ varies.
  • Pre-treatment pCR prediction from baseline DCE-MRI alone, in this benchmark, is statistically indistinguishable from random for all but one of fourteen submissions; relying on such models clinically would be unsafe.
  • The absence of a link between segmentation quality and pCR performance suggests the limiting factor is response representation, not tumor delineation.
  • Future progress depends on redefining the task: continuous targets (tumor shrinkage, residual burden, time-to-response), longitudinal imaging, and complementary modalities are more promising than simply scaling binary classifiers.

Where Pith is reading between the lines

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

  • If the external test set's pCR prevalence and tumor mix match routine clinical populations, the near-random result likely extends to other settings; a useful check would be to run the same public training data against other independent cohorts to see whether any model rises above chance.
  • The mean-gap fairness metric used here can be gamed by degrading the best-performing subgroup without helping anyone; reporting per-subgroup error distributions, not just gaps, is a natural extension a reader should request before trusting fairness scores.
  • The conclusion that pre-treatment imaging alone is insufficient points to a testable clinical design: compare pCR prediction from baseline MRI against prediction from early on-treatment MRI after one or two cycles; if on-treatment imaging restores discriminative accuracy, the missing signal is temporal rather than entirely biological.
  • The observed ranking instability under λ suggests that challenge leaderboards should publish results along the full accuracy–fairness trade-off curve, not a single composite rank, so downstream users can choose the operating point matched to their clinical risk tolerance.

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

3 major / 5 minor

Summary. The paper introduces the MAMA-MIA Challenge, a benchmark for breast DCE-MRI that jointly evaluates primary tumor segmentation (Task 1) and prediction of pathologic complete response (Task 2) from pre-treatment imaging. Training uses 1,506 patients from the public MAMA-MIA dataset (US multi-center); validation/test uses 574 patients from three European centers (GUMED, KAUNO, HCB), with a held-out test of 516. Submissions are scored by a composite S=(1−λ)S_p+λS_f with λ=0.5, combining task performance with subgroup fairness across age, menopausal status, and density. Twenty-six teams participated. Task 1 shows top teams improving over an nnU-Net baseline in both DSC and fairness. Task 2 shows near-random performance overall, with one team's balanced accuracy differing from a random baseline (p=0.014, unadjusted) and another marginal (p=0.063); calibration is poor. The authors conclude that pre-treatment DCE-MRI alone is unreliable for pCR prediction in multi-center settings and discuss implications for fairness-aware benchmarking.

Significance. The manuscript provides a substantial community resource: a standardized external test set, an evaluation pipeline, public code/data, and a fairness-aware scoring framework with λ-sensitivity analysis. The segmentation analysis is careful, including failure modes and per-center/tumor-size stratification. The Task 2 negative result, if robust, is an important evidence point against single-center pCR claims. The main weaknesses are statistical: the central Task 2 conclusion is not supported by per-center analysis, and the fairness scores that drive rankings lack uncertainty quantification. These are fixable with additional analyses, making the paper suitable for major revision.

major comments (3)
  1. [V-B, Table V, III-A.2] The central negative claim—'pretreatment prediction from baseline DCE-MRI alone proved unreliable in the multi-center setting'—is supported only by pooled balanced accuracy across the 516-case test set. The paper itself shows strong center effects in segmentation (Fig. 2) and differing tumor-size distributions across centers (Fig. 3). GUMED contributes 30 cases, KAUNO 232, HCB 312, and pCR prevalence may differ across sites. No center-stratified Task 2 results (balanced accuracy, pCR prevalence, calibration, significance) are reported. The observed near-random pooled performance is equally consistent with a genuinely absent imaging signal or with a signal that is present in some centers but diluted by between-center heterogeneity in NAC protocols, pCR definitions, or acquisition parameters. Please add per-center results or soften the conclusion to the pooled cohort.
  2. [III-E, Table V] Fairness scores are point estimates without confidence intervals, despite being a principal ranking component. For Task 2, the combined score at λ=0.5 is dominated by the fairness term: pimed-lab ranks first despite a performance score of 0.4964 (below baseline 0.5042) because its fairness score is much lower. GUMED contains only 30 cases, and per-subgroup pCR-positive counts will be very small; the max-minus-min subgroup gap is highly sensitive to small samples. Moreover, the paper acknowledges in Section V-C that mean-gap fairness can be 'gamed' by degrading the best-performing subgroup. Report bootstrap/permutation CIs for fairness gaps, or at least give subgroup sample sizes and pCR counts, and discuss the stability of the resulting rankings.
  3. [V-B, Table V] The statement 'Only one team achieved performance significantly different from random prediction (p=0.014)' uses unadjusted per-team p-values. With 14 Task 2 submissions, a Bonferroni threshold of 0.05/14≈0.0036 would render PM's p=0.014 non-significant, and FME's marginal p=0.063 clearly so. Since the near-random result is a load-bearing conclusion, report multiple-comparison-corrected p-values (or justify why correction is unnecessary). This does not necessarily change the negative conclusion, but it makes the 'only one team' formulation accurate.
minor comments (5)
  1. [III-E.2] HD_max is stated to be 'derived from the maximum distance observed in the automatic segmentation baseline'; this value is not reproducible as written. State the exact value and the baseline definition used to derive it.
  2. [III-A.2] The text gives a validation/test split of 58/516, but Table I presents only the combined 574 numbers. Clarify how the split aligns with Table I and whether each center is stratified identically in the validation and test portions.
  3. [Table IV] The baseline row is marked with an asterisk, but the footnote states that asterisked values are not statistically different from the baseline. This is circular; clarify that the baseline is the comparator and not itself tested.
  4. [V-B, Table V] Specify whether the p-values are one- or two-sided and describe the exact simulation procedure (e.g., 1000 random simulations) behind the comparisons in Table V.
  5. [III-F, V-B] The Task 2 baseline is a random classifier, whose expected balanced accuracy (0.5) coincides with that of a majority-class predictor. Reporting a majority-class baseline as well would help readers interpret the clinical meaning of 'above random'.

Circularity Check

0 steps flagged

No significant circularity: the paper's claims are empirical evaluations against an external test set, not derivations from the benchmark definition.

full rationale

The paper's central claims—robust multi-center segmentation and unreliable pretreatment pCR prediction—are empirical results obtained from submissions evaluated on a held-out external European test set (516 cases from GUMED, KAUNO, and HCB). The scoring equations (1)–(7) define metrics and fairness criteria but do not encode any particular outcome; they are evaluation instruments, not fitted predictors. The Task 2 baseline is explicitly a random classifier, and the finding that most methods do not significantly beat random is a statistical observation, not a consequence of the scoring definition. The Task 1 baseline nnU-Net is a reference method, not a parameter fitted to reproduce the leaderboard. The constant HD_max=150 mm is a normalization choice derived from baseline segmentation errors; it affects metric scaling but does not determine whether any method outperforms another. Self-citations, most prominently the MAMA-MIA dataset publication [6], supply the training data and annotation provenance; they do not serve as the argument for the paper's conclusions. The absence of per-center Task 2 significance testing is a legitimate external-validity concern, but that is a question of inferential strength, not circularity. No load-bearing equation or cited result reduces to its own input, so the derivation chain is self-contained with respect to circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No invented entities are proposed. The free parameters are evaluation-design constants (lambda, HD_max) and analytical thresholds; they are choices by the organizers rather than physical parameters. The axioms are mostly standard metric definitions plus two design decisions the paper itself flags as normative.

free parameters (4)
  • lambda (fairness weight) = 0.5
    Fairness weight in Eq. (1), chosen by organizers; determines the trade-off and substantially changes rankings as shown in Fig. 8.
  • HD_max = 150 mm
    Normalization constant in Eq. (4c), 'derived from the maximum distance observed in the automatic segmentation baseline'; directly affects NormHD and thus the segmentation performance score.
  • Tumor size thresholds = 20th and 80th percentiles of tumor volumes
    Used in Fig. 3 to define small/moderate/large categories; analysis-only, not part of the benchmark score, but shapes the reported size-dependence conclusions.
  • Age subgroup bins = <40, 41-50, 51-60, 61-70, >71
    Subgroup definition in Eq. (3a); the chosen discretization affects the fairness score and the subgroup analysis.
axioms (4)
  • standard math Dice similarity, normalized Hausdorff distance, balanced accuracy, and equalized-odds disparity are valid performance and fairness measures for these tasks.
    Used without derivation in Section III.E (Eqs. 4-7); standard definitions, but their selection shapes the benchmark outcomes.
  • domain assumption The external test cohort from GUMED, KAUNO, and HCB is representative of the cross-institutional generalization target.
    Section III.A.2 and Table I define the cohort; the paper's generalizability claims depend on this cohort standing in for unseen clinical sites.
  • ad hoc to paper Fairness is adequately summarized by the maximum-minus-minimum subgroup mean gap, averaged over tasks and variables.
    Eqs. (5) and (7) define S_f; the paper itself notes in Section V.C that mean-based gaps can hide variance and be reduced by degrading privileged groups.
  • ad hoc to paper Constant (all-zero/all-one) pCR submissions are invalid and can be excluded from evaluation.
    Section III.E.3 states this exclusion; it removes degenerate classifiers that would otherwise appear perfectly fair.

pith-pipeline@v1.3.0-alltime-deepseek · 19843 in / 13315 out tokens · 128271 ms · 2026-08-02T19:39:25.531790+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of The MAMA-MIA Challenge: Advancing Generalizability and Fairness in Breast MRI Tumor Segmentation and Treatment Response Prediction." pith.science (2026). https://pith.science/paper/ZSV56MII

@misc{pith2026260301250,
  author       = {Pith},
  title        = {Pith review of: The MAMA-MIA Challenge: Advancing Generalizability and Fairness in Breast MRI Tumor Segmentation and Treatment Response Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZSV56MII}},
  note         = {Machine review of arXiv:2603.01250}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Breast cancer is the most frequently diagnosed malignancy among women worldwide and a leading cause of cancer-related mortality. Dynamic contrast-enhanced magnetic resonance imaging plays a central role in tumor characterization and treatment monitoring, particularly in patients receiving neoadjuvant chemotherapy. However, existing artificial intelligence models for breast magnetic resonance imaging are typically developed and evaluated using heterogeneous datasets, study populations, and assessment protocols, making direct comparison difficult and limiting understanding of model robustness across institutions and clinically relevant patient subgroups. The MAMA-MIA Challenge was designed to address these challenges by providing a standardized benchmark for the joint evaluation of primary tumor segmentation and prediction of pathologic complete response using pre-treatment magnetic resonance imaging only. The training cohort comprised 1,506 patients from multiple institutions in the United States, while evaluation was conducted on an external test set of 574 patients from three independent European centers to assess cross-continental and cross-institutional generalization. A unified scoring framework combined predictive performance with subgroup consistency across age, menopausal status, and breast density. Twenty-six international teams participated in the final evaluation phase. Results demonstrate substantial performance variability under a common external evaluation framework and reveal trade-offs between overall accuracy and subgroup fairness. The challenge provides standardized datasets, evaluation protocols, and public resources to promote the development of robust and equitable artificial intelligence systems for breast cancer imaging.

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

65 extracted references · 2 canonical work pages

  1. [1]

    Breast cancer: Global patterns of incidence, mortality, and survival,

    World Health Organization, “Breast cancer: Global patterns of incidence, mortality, and survival,” https://www.who.int/news-room/fact-sheets/ detail/breast-cancer, 2024, accessed: 2025-04-27

    2024

  2. [2]

    Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries,

    H. Sunget al., “Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries,” CA: a cancer journal for clinicians, vol. 71, no. 3, pp. 209–249, 2021

    2020

  3. [3]

    Breast MRI: state of the art,

    R. M. Mann, N. Cho, and L. Moy, “Breast MRI: state of the art,” Radiology, vol. 292, no. 3, pp. 520–536, 2019

    2019

  4. [4]

    Locally advanced breast cancer: MR imaging for prediction of response to neoadjuvant chemotherapy—results from ACRIN 6657/I-SPY trial,

    N. Hylton and et al., “Locally advanced breast cancer: MR imaging for prediction of response to neoadjuvant chemotherapy—results from ACRIN 6657/I-SPY trial,”Radiology, vol. 279, no. 1, pp. 44–55, 2016

    2016

  5. [5]

    Rele- vance of pathological complete response after neoadjuvant therapy for breast cancer,

    A. Pennisi, T. Kieber-Emmons, I. Makhoul, and L. Hutchins, “Rele- vance of pathological complete response after neoadjuvant therapy for breast cancer,”Breast cancer: basic and clinical research, vol. 10, pp. BCBCR–S33 163, 2016

    2016

  6. [6]

    A large-scale multicenter breast cancer dce-mri benchmark dataset with expert segmentations,

    L. Garruchoet al., “A large-scale multicenter breast cancer dce-mri benchmark dataset with expert segmentations,”Scientific Data, vol. 12, no. 1, p. 453, 2025

    2025

  7. [7]

    A robust and efficient ai assistant for breast tumor segmentation from dce-mri via a spatial-temporal framework,

    J. Zhanget al., “A robust and efficient ai assistant for breast tumor segmentation from dce-mri via a spatial-temporal framework,”Patterns, vol. 4, no. 9, 2023

    2023

  8. [8]

    Deep learning prediction of pathologic complete response in breast cancer using mri and other clinical data: a systematic review,

    N. Khan, R. Adam, P. Huang, T. Maldjian, and T. Q. Duong, “Deep learning prediction of pathologic complete response in breast cancer using mri and other clinical data: a systematic review,”Tomography, vol. 8, no. 6, pp. 2784–2795, 2022

    2022

  9. [9]

    Algorithmic fairness in artificial intelligence for medicine and healthcare,

    R. J. Chenet al., “Algorithmic fairness in artificial intelligence for medicine and healthcare,”Nature biomedical engineering, vol. 7, no. 6, pp. 719–742, 2023

    2023

  10. [10]

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

    K. Lekadiret al., “Future-ai: international consensus guideline for trustworthy and deployable artificial intelligence in healthcare,”bmj, vol. 388, 2025

    2025

  11. [11]

    Breast multiparametric mri for prediction of neoadjuvant chemotherapy response in breast cancer: the bmmr2 challenge,

    W. Liet al., “Breast multiparametric mri for prediction of neoadjuvant chemotherapy response in breast cancer: the bmmr2 challenge,”Radi- ology: Imaging Cancer, vol. 6, no. 1, p. e230033, 2024

    2024

  12. [12]

    The limits of fair medical imaging ai in real-world generalization,

    Y . Yang, H. Zhang, J. W. Gichoya, D. Katabi, and M. Ghassemi, “The limits of fair medical imaging ai in real-world generalization,”Nature Medicine, vol. 30, no. 10, pp. 2838–2848, 2024

    2024

  13. [13]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inInternational Conference on Medical image computing and computer-assisted intervention. Springer, 2015, pp. 234–241

    2015

  14. [14]

    nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,

    F. Isensee, P. F. Jaeger, S. A. A. Kohl, J. Petersen, and K. H. Maier-Hein, “nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,”Nature Methods, vol. 18, no. 2, pp. 203–211,

  15. [15]

    3d mri brain tumor segmentation using autoencoder reg- ularization,

    A. Myronenko, “3d mri brain tumor segmentation using autoencoder reg- ularization,” inInternational MICCAI brainlesion workshop. Springer, 2018, pp. 311–320

    2018

  16. [16]

    Generalizing to unseen domains: A survey on domain generalization,

    J. Wanget al., “Generalizing to unseen domains: A survey on domain generalization,”IEEE Transactions on Knowledge and Data Engineer- ing, pp. 1–1, 2022

    2022

  17. [17]

    Breast mri background parenchymal enhance- ment categorization using deep learning: Outperforming the radiologist,

    S. Eskreis-Winkleret al., “Breast mri background parenchymal enhance- ment categorization using deep learning: Outperforming the radiologist,” Journal of Magnetic Resonance Imaging, vol. 56, no. 4, pp. 1068–1076, Oct 2022, epub 2022 Feb 15

    2022

  18. [18]

    Mammographic density and ageing: a collaborative pooled analysis of cross-sectional data from 22 countries worldwide,

    A. Burtonet al., “Mammographic density and ageing: a collaborative pooled analysis of cross-sectional data from 22 countries worldwide,” PLoS medicine, vol. 14, no. 6, p. e1002335, 2017

    2017

  19. [19]

    A machine learning model that classifies breast can- cer pathologic complete response on mri post-neoadjuvant chemother- apy,

    E. J. Suttonet al., “A machine learning model that classifies breast can- cer pathologic complete response on mri post-neoadjuvant chemother- apy,”Breast Cancer Research, vol. 22, pp. 1–11, 2020

    2020

  20. [20]

    Comparison of radiomics-based machine-learning classifiers for the pretreatment pre- diction of pathologic complete response to neoadjuvant therapy in breast cancer,

    X. Li, C. Li, H. Wang, L. Jiang, and M. Chen, “Comparison of radiomics-based machine-learning classifiers for the pretreatment pre- diction of pathologic complete response to neoadjuvant therapy in breast cancer,”PeerJ, vol. 12, p. e17683, 2024

    2024

  21. [21]

    Pretreatment dce-mri-based deep learning outperforms radiomics analysis in predicting pathologic complete response to neoad- juvant chemotherapy in breast cancer,

    Y . Penget al., “Pretreatment dce-mri-based deep learning outperforms radiomics analysis in predicting pathologic complete response to neoad- juvant chemotherapy in breast cancer,”Frontiers in oncology, vol. 12, p. 846775, 2022. 14 IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. XX, NO. XX, XXXX 2020

    2022

  22. [22]

    Deep-learning based discrimination of pathologic complete response using mri in her2- positive and triple-negative breast cancer,

    S.-Y . Kim, J. Lee, N. Cho, and Y .-G. Kim, “Deep-learning based discrimination of pathologic complete response using mri in her2- positive and triple-negative breast cancer,”Scientific Reports, vol. 14, no. 1, p. 23065, 2024

    2024

  23. [23]

    Machine learning with magnetic reso- nance imaging for prediction of response to neoadjuvant chemotherapy in breast cancer: A systematic review and meta-analysis,

    X. Liang, X. Yu, and T. Gao, “Machine learning with magnetic reso- nance imaging for prediction of response to neoadjuvant chemotherapy in breast cancer: A systematic review and meta-analysis,”European Journal of Radiology, vol. 150, p. 110247, 2022

    2022

  24. [24]

    Domain generalization in deep learning based mass detection in mam- mography: A large-scale multi-center study,

    L. Garrucho, K. Kushibar, S. Jouide, O. Diaz, L. Igual, and K. Lekadir, “Domain generalization in deep learning based mass detection in mam- mography: A large-scale multi-center study,”Artificial Intelligence in Medicine, vol. 132, p. 102386, 2022

    2022

  25. [25]

    Identifying the best machine learning algorithms for brain tumor segmentation, progression assessment, and overall survival prediction in the brats challenge,

    S. Bakaset al., “Identifying the best machine learning algorithms for brain tumor segmentation, progression assessment, and overall survival prediction in the brats challenge,”arXiv preprint arXiv:1811.02629, 2018

    Pith/arXiv arXiv 2018

  26. [26]

    Multi-centre, multi-vendor and multi-disease cardiac segmentation: the m&ms challenge,

    V . Campelloet al., “Multi-centre, multi-vendor and multi-disease cardiac segmentation: the m&ms challenge,”IEEE Transactions on Medical Imaging, vol. 40, no. 12, pp. 3543–3554, 2021

    2021

  27. [27]

    Crossmoda 2021 challenge: Benchmark of cross- modality domain adaptation techniques for vestibular schwannoma and cochlea segmentation,

    R. Dorentet al., “Crossmoda 2021 challenge: Benchmark of cross- modality domain adaptation techniques for vestibular schwannoma and cochlea segmentation,”Medical Image Analysis, vol. 83, p. 102628, 2023

    2021

  28. [28]

    Toward fairness in artificial intelligence for medical image analysis: identification and mitigation of potential biases in the roadmap from data collection to model deployment,

    K. Drukkeret al., “Toward fairness in artificial intelligence for medical image analysis: identification and mitigation of potential biases in the roadmap from data collection to model deployment,”Journal of Medical Imaging, vol. 10, no. 6, pp. 061 104–061 104, 2023

    2023

  29. [29]

    Auditing unfair biases in cnn-based diagnosis of alzheimer’s disease,

    V . N. Danget al., “Auditing unfair biases in cnn-based diagnosis of alzheimer’s disease,” inWorkshop on Clinical Image-Based Procedures. Springer, 2023, pp. 172–182

    2023

  30. [30]

    Breast cancer: Global pat- terns of incidence, mortality, and trends

    E. T. Sedeta, B. Jobre, and B. Avezbakiyev, “Breast cancer: Global pat- terns of incidence, mortality, and trends.”Journal of Clinical Oncology, vol. 41, no. 16_suppl, pp. 10 528–10 528, 2023

    2023

  31. [31]

    The cancer imaging archive (tcia): maintaining and operating a public information repository,

    K. Clarket al., “The cancer imaging archive (tcia): maintaining and operating a public information repository,”Journal of digital imaging, vol. 26, no. 6, pp. 1045–1057, 2013

    2013

  32. [32]

    Multicenter breast DCE-MRI data and segmentations from patients in the I-SPY 1/ACRIN 6657 trials,

    D. Newittet al., “Multicenter breast DCE-MRI data and segmentations from patients in the I-SPY 1/ACRIN 6657 trials,”The Cancer Imaging Archive, 2016. [Online]. Available: https://doi.org/10.7937/K9/TCIA. 2016.HdHpgJLK

    work page doi:10.7937/k9/tcia 2016

  33. [33]

    I-SPY 2 breast dynamic contrast enhanced MRI trial (version 1) [data set],

    W. Liet al., “I-SPY 2 breast dynamic contrast enhanced MRI trial (version 1) [data set],”The Cancer Imaging Archive, 2022. [Online]. Available: https://doi.org/10.7937/TCIA.D8Z0-9T85

    doi:10.7937/tcia.d8z0-9t85 2022

  34. [34]

    Single site breast DCE-MRI data and segmentations from patients undergoing neoadjuvant chemotherapy (version 3) [data set],

    D. Newitt and N. Hylton, “Single site breast DCE-MRI data and segmentations from patients undergoing neoadjuvant chemotherapy (version 3) [data set],”The Cancer Imaging Archive, 2016. [Online]. Available: https://doi.org/10.7937/K9/TCIA.2016.QHsyhJKy

    doi:10.7937/k9/tcia.2016.qhsyhjky 2016

  35. [35]

    Dynamic contrast-enhanced magnetic resonance images of breast cancer patients with tumor locations [data set],

    A. Sahaet al., “Dynamic contrast-enhanced magnetic resonance images of breast cancer patients with tumor locations [data set],” The Cancer Imaging Archive, 2021. [Online]. Available: https: //doi.org/10.7937/TCIA.e3sv-re93

    doi:10.7937/tcia.e3sv-re93 2021

  36. [36]

    Pycad: A python library for medical imaging analysis,

    M. Amine, “Pycad: A python library for medical imaging analysis,” https://github.com/amine0110/pycad, 2023, gitHub repository

    2023

  37. [37]

    Codabench: Flexible, easy-to-use, and reproducible meta- benchmark platform,

    Z. Xuet al., “Codabench: Flexible, easy-to-use, and reproducible meta- benchmark platform,”Patterns, vol. 3, no. 7, 2022

    2022

  38. [38]

    J. C. Geeet al.,Medical Image Computing and Computer Assisted Intervention–MICCAI 2025: 28th International Conference, Daejeon, South Korea, September 23–27, 2025, Proceedings, Part V. Springer Nature, 2025

    2025

  39. [39]

    Zhanget al., Eds.,Artificial Intelligence and Imaging for Diagnostic and Treatment Challenges in Breast Care, ser

    T. Zhanget al., Eds.,Artificial Intelligence and Imaging for Diagnostic and Treatment Challenges in Breast Care, ser. Lecture Notes in Computer Science, vol. 16142. Cham: Springer, 2026. [Online]. Available: https://link.springer.com/book/10.1007/978-3-032-05559-0

    work page doi:10.1007/978-3-032-05559-0 2026

  40. [40]

    nnu-net revisited: A call for rigorous validation in 3d medical image segmentation,

    F. Isenseeet al., “nnu-net revisited: A call for rigorous validation in 3d medical image segmentation,” inMedical Image Computing and Computer Assisted Intervention – MICCAI 2024. Cham: Springer Nature Switzerland, 2024, pp. 488–498

    2024

  41. [41]

    On tackling domain shift in breast MRI using only publicly-available data: Reproducible breast cancer segmentation and pCR prediction,

    J. Kächele, D. Bounias, A. Ertl, and K. Maier-Hein, “On tackling domain shift in breast MRI using only publicly-available data: Reproducible breast cancer segmentation and pCR prediction,” inLecture Notes in Computer Science, ser. Lecture notes in computer science. Cham: Springer Nature Switzerland, 2026, pp. 310–319

    2026

  42. [42]

    Can we teach ai to understand breast tumour behaviour? our mama-mia challenge journey,

    H. Awwad, J. C. Vilanova, and R. Martí, “Can we teach ai to understand breast tumour behaviour? our mama-mia challenge journey,” inArtificial Intelligence and Imaging for Diagnostic and Treatment Challenges in Breast Care, T. Zhanget al., Eds. Cham: Springer Nature Switzerland, 2026, pp. 248–257

    2026

  43. [43]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  44. [44]

    A generalizable 3d framework and model for self- supervised learning in medical imaging,

    T. Xuet al., “A generalizable 3d framework and model for self- supervised learning in medical imaging,” 2025. [Online]. Available: https://arxiv.org/abs/2501.11755

    Pith/arXiv arXiv 2025

  45. [45]

    Divergence-aware training with automatic subgroup mitigation for breast tumor segmentation,

    E. Poetaet al., “Divergence-aware training with automatic subgroup mitigation for breast tumor segmentation,” inDeep Breast Workshop on AI and Imaging for Diagnostic and Treatment Challenges in Breast Care. Springer, 2025, pp. 52–62

    2025

  46. [46]

    Looking for trouble: Analyzing classifier behavior via pattern divergence,

    E. Pastor, L. de Alfaro, and E. Baralis, “Looking for trouble: Analyzing classifier behavior via pattern divergence,” inProceedings of the 2021 International Conference on Management of Data, ser. SIGMOD ’21. New York, NY , USA: Association for Computing Machinery, 2021, p. 1400–1412. [Online]. Available: https://doi.org/10.1145/3448016. 3457284

    doi:10.1145/3448016 2021

  47. [47]

    Mitigating sub- group disparities in speech models: A divergence-aware dual strategy,

    A. Koudounas, E. Pastor, L. de Alfaro, and E. Baralis, “Mitigating sub- group disparities in speech models: A divergence-aware dual strategy,” IEEE Transactions on Audio, Speech and Language Processing, vol. 33, pp. 883–895, 2025

    2025

  48. [48]

    A simple framework for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A simple framework for contrastive learning of visual representations,” 2020. [Online]. Available: https://arxiv.org/abs/2002.05709

    Pith/arXiv arXiv 2020

  49. [49]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” 2015. [Online]. Available: https://arxiv.org/abs/1512.03385

    Pith/arXiv arXiv 2015

  50. [50]

    A closer look at spatiotemporal convolutions for action recognition,

    D. Tran, H. Wang, L. Torresani, J. Ray, Y . LeCun, and M. Paluri, “A closer look at spatiotemporal convolutions for action recognition,”

  51. [51]

    (2016) TorchVision: PyTorch’s computer vision library

    TorchVision maintainers and contributors. (2016) TorchVision: PyTorch’s computer vision library. GitHub repository. [Online]. Available: https://github.com/pytorch/vision

    2016

  52. [52]

    Tune: A research platform for distributed model selection and training,

    R. Liaw, E. Liang, R. Nishihara, P. Moritz, J. E. Gonzalez, and I. Stoica, “Tune: A research platform for distributed model selection and training,” 2018. [Online]. Available: https://arxiv.org/abs/1807.05118

    Pith/arXiv arXiv 2018

  53. [53]

    Xgboost: A scalable tree boosting system,

    T. Chen and C. Guestrin, “Xgboost: A scalable tree boosting system,” inProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ser. KDD ’16. ACM, Aug. 2016, p. 785–794. [Online]. Available: http: //dx.doi.org/10.1145/2939672.2939785

    arXiv 2016

  54. [54]

    Caballo, W

    M. Caballo, W. B. Sanderink, L. Han, Y . Gao, A. Athanasiou, and R. M. Mann, “Four-dimensional machine learning radiomics for the pre- treatment assessment of breast cancer pathologic complete response to neoadjuvant chemotherapy in dynamic contrast-enhanced mri,”Journal of Magnetic Resonance Imaging, vol. 57, no. 1, pp. 97–110, 2023

    2023

  55. [55]

    Temporal representation learning of phenotype trajectories for pcr prediction in breast cancer,

    I. Janí ˇckováet al., “Temporal representation learning of phenotype trajectories for pcr prediction in breast cancer,” inInternational Confer- ence on Medical Image Computing and Computer-Assisted Intervention. Springer, 2025, pp. 606–615

    2025

  56. [56]

    Predicting neoadjuvant chemotherapy treatment response in hormone-receptor-positive/her2- negative breast cancer–results from the swedish scan-b population-based cohort,

    N. Loman, H. Saghir, and S. Kimbung, “Predicting neoadjuvant chemotherapy treatment response in hormone-receptor-positive/her2- negative breast cancer–results from the swedish scan-b population-based cohort,”Acta Oncologica, vol. 64, p. 44201, 2025

    2025

  57. [57]

    Association of menopausal status, expression of progesterone receptor and ki67 to the clinical response to neoadjuvant chemotherapy in luminal breast cancer,

    L. R. da Silva, R. F. Vargas, J. Y . Shinzato, S. F. M. Derchain, S. Ramalho, and L. C. Zeferino, “Association of menopausal status, expression of progesterone receptor and ki67 to the clinical response to neoadjuvant chemotherapy in luminal breast cancer,”Revista Brasileira de Ginecologia e Obstetrícia/RBGO Gynecology and Obstetrics, vol. 41, no. 12, pp....

    2019

  58. [58]

    Fairness-aware data augmentation for cardiac mri using text-conditioned diffusion models,

    G. Skorupkoet al., “Fairness-aware data augmentation for cardiac mri using text-conditioned diffusion models,” inMICCAI Workshop on Fairness of AI in Medical Imaging. Springer, 2025, pp. 63–73

    2025

  59. [59]

    High-resolution synthesis of high-density breast mammograms: Application to improved fairness in deep learning based mass detection,

    L. Garruchoet al., “High-resolution synthesis of high-density breast mammograms: Application to improved fairness in deep learning based mass detection,”Frontiers in oncology, vol. 12, p. 1044496, 2023

    2023

  60. [60]

    The predictive value of ki-67 before neoadjuvant chemotherapy for breast cancer: a systematic review and meta-analysis,

    X. Chenet al., “The predictive value of ki-67 before neoadjuvant chemotherapy for breast cancer: a systematic review and meta-analysis,” Future oncology, vol. 13, no. 9, p. 843, 2017

    2017

  61. [61]

    Ai-driven mri biomarker for triple-class her2 expression classification in breast cancer: a large-scale multicenter study,

    C. Wonget al., “Ai-driven mri biomarker for triple-class her2 expression classification in breast cancer: a large-scale multicenter study,”Breast Cancer Research, vol. 27, no. 1, p. 166, 2025

    2025

  62. [62]

    Highly accurate response prediction in high-risk early breast cancer patients using a biophysical simulation platform,

    F. M. Howardet al., “Highly accurate response prediction in high-risk early breast cancer patients using a biophysical simulation platform,” Breast Cancer Research and Treatment, vol. 196, no. 1, pp. 57–66, 2022

    2022

  63. [63]

    Federated nnu-net for privacy-preserving medical image segmentation,

    G. Skorupkoet al., “Federated nnu-net for privacy-preserving medical image segmentation,”Scientific Reports, vol. 15, no. 1, p. 38312, 2025

    2025

  64. [2018]

    Available: https://arxiv.org/abs/1711.11248

    [Online]. Available: https://arxiv.org/abs/1711.11248

    Pith/arXiv arXiv

  65. [2021]

    Available: https://doi.org/10.1038/s41592-020-01008-z

    [Online]. Available: https://doi.org/10.1038/s41592-020-01008-z

    doi:10.1038/s41592-020-01008-z