Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

MeDi: Metadata-Guided Diffusion Models for Mitigating Biases in Tumor Classification

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Conditioning diffusion models on medical center metadata improves tumor classification under distribution shift.

desk verdict MeDi shows a plausible data-centric debiasing mechanism, but the paper never tests the fidelity of the exact synthetic recombinations that drive its downstream gains. read the letter →

arxiv 2506.17140 v1 pith:R33YDG22 submitted 2025-06-20 eess.IV cs.AIcs.CV

David Jacob Drexlin , Jonas Dippel , Julius Hense , Niklas Preni{ss}l , Grégoire Montavon , Frederick Klauschen , Klaus-Robert Müller This is my paper
classification eess.IVcs.AIcs.CV
keywords diffusionmodelshistopathologymetadataconditioningshortcutlearningdataaugmentationtumorclassificationdomainshiftTCGA
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 proposes MeDi, a diffusion model for histopathology images that is conditioned not only on tumor class but also on metadata such as the medical center (tissue source site). The authors argue that this extra conditioning lets the generator reproduce the staining and technical style of specific centers, producing higher-fidelity synthetic patches (average FID 37.73 vs 50.65) and enabling targeted augmentation of underrepresented class–site combinations. They show that adding one synthetic MeDi image per real sample improves a linear classifier on frozen foundation-model embeddings for NSCLC subtyping (81.37 vs 74.31 without synthetic data) and uterine subtyping (74.86 vs 67.73), while the class-only diffusion baseline helps less. If true, MeDi provides a data-centric way to reduce shortcut learning and hospital batch effects in computational pathology.

What carries the argument

The central object is a metadata-conditioned diffusion UNet. A class-label embedding and a learnable embedding for each metadata attribute (here, tissue source site) are concatenated into a conditioning vector matched to the timestep embedding dimension, then added to the timestep embedding at each residual block. During inference the user can condition on any class–metadata combination, including combinations absent from training, allowing the model to interpolate or extrapolate within metadata space. This is what lets MeDi fill missing subpopulations and break the spurious correlation between tissue source site and tumor class.

What would settle it

Train a weakly supervised whole-slide classifier with and without MeDi augmentation on the same TCGA-UT tasks and evaluate on an external cohort of medical centers, comparing balanced accuracy for sites unseen in both training and generation; if the augmented model is not better than the unaugmented one, the claim that MeDi mitigates tumor classification bias in practice fails.

Watch

Extended reading notes

Core claim

The central claim is that explicitly conditioning a diffusion model on tissue source site in addition to the cancer type yields two benefits: generated histopathology patches more closely match the real training distribution, and downstream tumor-subtype classifiers generalize better to medical centers the diffusion model never saw. On 32 cancer types from TCGA, metadata conditioning lowers average Fréchet Inception Distance from 50.65 to 37.73, with 23 of 32 cancers improving. In subpopulation-shift experiments, augmenting training with one synthetic MeDi patch per real patch improves balanced accuracy on unseen tissue source sites for NSCLC (81.37% vs 74.31% without synthetic data) and uterine cancer (74.86% vs 67.73%), while class-only diffusion gives intermediate gains. MeDi is offered as a proof-of-concept that generative augmentation can complement foundation models in mitigating Clever Hans effects and demographic or hospital biases.

Load-bearing premise

The load-bearing premise is that a linear classifier on frozen UNI embeddings, trained on 20 samples per class with one synthetic image per real sample, is representative of how tumor-classification models perform in real clinical deployment, where weakly supervised and fine-tuned models dominate.

Editorial extensions

If this is right

  • If MeDi's improvement is real, targeted synthetic augmentation is a viable data-centric complement to foundation-model embeddings for subtyping tumors when some hospitals are absent from training.
  • The large FID drop (50.65 to 37.73) indicates that conditioning on tissue source site reduces the generator's tendency to collapse onto the dominant stain or style mode of each cancer class.
  • MeDi's ability to generate combinations of class and site that were never seen in training gives a practical tool to rebalance skewed class–metadata distributions.
  • Because tissue source site is only one of several available metadata attributes, the same mechanism extends in principle to scanner type, patient demographics, and staining protocol, though the paper only tests tissue source site.

Reading between the lines

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

  • Editorial inference: the reported gains come from a linear probe on frozen UNI embeddings with 20 training samples per class; whether fine-tuning the foundation model or using weakly supervised whole-slide training preserves the benefit is untested and is the main transfer question.
  • Editorial inference: the FID improvement may largely reflect matching site-specific stain statistics; a direct comparison against stain normalization or CycleGAN normalization would clarify whether the gain is generative or purely color-related.
  • Editorial inference: a natural extension is to use MeDi samples as a privacy-preserving surrogate for underrepresented hospital sites, but the paper does not address privacy or data-sharing constraints.
  • Editorial inference: testing with classifier-free guidance or latent diffusion could show whether the metadata-conditioning gain scales with model capacity.
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 / 6 minor

Summary. The paper proposes MeDi, a metadata-guided diffusion model for histopathology patches, in which a class label (cancer type) and a metadata attribute (tissue source site, TSS) are embedded and added to the timestep embedding of a standard UNet. On the TCGA-UT dataset, the authors compare a class-only (CLS) diffusion model with MeDi on two fronts: (i) fidelity, measured by FID between real and synthetic patches sampled according to the training distribution, where MeDi reports an average FID of 37.73 versus 50.65 for CLS; and (ii) downstream tumor-subtype classification, where a linear classifier on frozen UNI embeddings is trained with additional synthetic patches on NSCLC, RCC, and uterine subtyping tasks under a subpopulation-shift holdout. The paper reports accuracy gains for MeDi over no-synthetic and CLS baselines on NSCLC and uterine tasks, and a roughly on-par or worse result on RCC. The authors position the work as a proof-of-concept for using metadata-conditioned generation to mitigate data biases in computational pathology.

Significance. If the central claims hold, MeDi would be a useful data-centric tool for reducing batch effects and shortcut learning in histopathology, with the practical advantage that it uses readily available metadata rather than requiring new annotations. The paper has several concrete strengths: the authors provide public code; the evaluation is causal and non-circular, with the diffusion training distribution excluding the test tissue source sites; the downstream evaluation uses a strong histopathology foundation model (UNI); and the experimental scope covers 32 cancer types and many medical centers. However, the manuscript currently overstates its results relative to the evidence: the fidelity advantage is not evaluated for the unseen class–TSS recombinations that drive the debiasing mechanism, the RCC row of Table 1 contradicts the blanket claim of improved downstream performance, and the training-set description in Section 4.4 is internally inconsistent. These gaps are local and addressable, but they are load-bearing for the paper's main narrative.

major comments (4)
  1. [4.3 and 4.4] The central debiasing mechanism relies on synthetic images for class–TSS combinations that are absent from diffusion training (e.g., generating (LUAD, TSS_B) when real data contain (LUAD, TSS_A) and (LUSC, TSS_B)), yet the fidelity evaluation in Section 4.3 samples only the training distribution, i.e., pairs that co-occur in training. No FID, class-preservation metric, or human evaluation is reported for the recombined pairs. If these recombinations are unrealistic or class-mixing, the downstream gains in Table 1 could be due to generic augmentation artifacts rather than genuine debiasing. The paper should either directly measure fidelity and class-consistency for the recombined pairs or explicitly restrict the fidelity claim to seen pairs and reframe the debiasing evidence accordingly.
  2. [Table 1] The abstract and conclusion state that MeDi 'enables improvements in performance for downstream classifiers,' but the RCC row of Table 1 shows MeDi overall balanced accuracy of 74.87, below the CLS-only baseline of 78.21 and only slightly above the no-synthetic baseline of 73.25. For TSS-averaged accuracy, MeDi (84.17) is also slightly below CLS (84.64). The paper's own text acknowledges only that MeDi is 'roughly on par' for RCC on TSS average, which is already a significant qualification. The blanket claims in the abstract and Section 5 should be tempered to specify the two tasks where the improvement holds, and the RCC result should be analyzed rather than passed over.
  3. [4.4] The description of the downstream training set is internally inconsistent. Section 4.4 states that 'In the training set, each cancer type is sourced from only one distinct TSS,' which contradicts the immediately preceding task descriptions (e.g., NSCLC 'both well represented across many centers') and the Section 4.1 holdout definition, which randomly excludes 30% of medical-center/race combinations. The sentence 'The test set contains data from all TSS that the diffusion models have not seen during training' also appears incompatible with 'we repeat the model training for each possible combination of TSS from the training set.' The exact construction of the training/test TSS sets and the 'all possible combinations' sampling procedure must be clarified, as the reproducibility and interpretation of Table 1 depend on it.
  4. [4.4 and 5] The downstream evidence is limited to a linear probe on frozen UNI embeddings trained with 20 samples per class and one synthetic image per real sample. This is a reasonable proof-of-concept, but the paper's general framing in the abstract and conclusion ('mitigating biases in tumor classification') goes beyond this setting. No experiment addresses fine-tuned foundation models or weakly supervised whole-slide classifiers, which are the settings used in actual clinical pipelines. The claims should either be explicitly scoped to low-data linear probing, or an additional experiment (even on one task) should show that the benefit transfers to a higher-capacity predictor.
minor comments (6)
  1. [4.1] There is a typo in the first sentence of Section 4.1: 'classification tasks„' should be 'classification tasks.'
  2. [4.4] In the text after Table 1, 'NCLSC' should be 'NSCLC.'
  3. [Table 1] The column header 'TSS A VG' should read 'TSS AVG.'
  4. [3.1] The conditioning mechanism adds z_cond directly to the timestep embedding (z_final = z_t + z_cond), which is an unusual design choice for a diffusion UNet; a sentence explaining why this was preferred over cross-attention or AdaGN would improve reproducibility.
  5. [Abstract and 4.3] The abstract says MeDi 'generates high-quality histopathology images for unseen subpopulations in TCGA,' but Section 4.3 evaluates FID only on training-distribution pairs; the claim about unseen subpopulations is not directly supported by any reported fidelity metric for held-out class–TSS combinations.
  6. [4.3] The average FID values (50.65 vs 37.73) are reported without variance, confidence intervals, or a significance test; given that FID has high sampling variability, the authors should report multiple generations or a bootstrap interval.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MeDi's training, sampling, and evaluation are causally separated, and no predicted quantity reduces to a fitted input or self-citation chain.

full rationale

The paper's central claims are not circular by construction. Fidelity is measured with FID between real and synthetic patches, and both the class-only (CLS) and metadata-conditioned (MeDi) diffusion models are trained on the same split with the same hyperparameters; the difference is the conditioning variable, not a quantity defined in terms of the outcome. The downstream generalization experiments are also causally clean: the diffusion model's training set excludes the test tissue source sites, and the augmentation policy (one synthetic example per real sample, uniform sampling over all class–TSS combinations) is fixed before evaluation, not fitted to the test performance. No parameter is calibrated on the target metric and then reported as a prediction. The paper's motivation cites several works that include the authors' own prior studies (e.g., Kauffmann et al. and Kömen et al.), but those citations are not load-bearing: the same premise is supported by external references (e.g., de Jong et al. and Vaidya et al.), and no result or uniqueness theorem is imported from a self-citation to force the paper's design. The abstract's phrase 'for unseen subpopulations' is broader than what Section 4.3 directly evaluates, since the FID setup samples seen class–TSS combinations; however, that is an evidence gap or framing issue, not a circular step. Taken as a whole, the derivation chain does not reduce to its own inputs by definition, so no circular step is identified.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on experimental hyperparameters rather than a mathematical derivation. The main hand-chosen settings are the 1:1 augmentation ratio, the 20-sample linear probe regime, and the 100-step DDIM sampler. Background assumptions include the validity of the diffusion objective, the use of UNI embeddings as the downstream feature space, and FID as a fidelity metric. No new physical entities are introduced.

free parameters (3)
  • Augmentation ratio = 1 synthetic image per real sample
    The downstream comparison in Table 1 uses exactly one synthetic patch per real training patch for both CLS and MeDi; the results are not shown to be stable across augmentation ratios, and the reported benefit may depend on this choice.
  • Linear probe training samples per class = 20
    Downstream classifiers are trained on UNI embeddings of 20 samples per class in Section 4.4; this low-data regime may favor augmentation and makes it unclear whether gains persist with more real data.
  • DDIM inference steps = 100
    All FID and downstream images are sampled with 100 DDIM steps; sample quality and therefore all reported metrics depend on this schedule.
assumptions (4)
  • standard math Standard diffusion training objective and DDIM sampling are correct and applicable to histopathology patches.
    Section 3.1 builds on the diffusers library UNet and uses DDIM with 100 steps; the paper does not re-derive the diffusion equations.
  • domain assumption UNI foundation model embeddings are a sufficient representation for tumor subtyping under hospital shift.
    Section 4.4 trains only a linear classifier on frozen UNI embeddings, so all downstream conclusions inherit the assumption that this embedding space captures the relevant class and batch variation.
  • domain assumption Tissue source site (TSS) is a valid proxy for hospital-related staining, scanner, and demographic batch effects.
    The whole MeDi mechanism conditions on TSS codes as the sole metadata attribute; if TSS labels do not track the actual nuisance variation, the method cannot help.
  • domain assumption FID computed with Inception features measures diagnostically relevant image fidelity in histopathology.
    Section 4.3 uses FID as the fidelity metric, but no pathologist evaluation or tissue-specific metric is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MeDi: Metadata-Guided Diffusion Models for Mitigating Biases in Tumor Classification." pith.science (2026). https://pith.science/paper/R33YDG22

@misc{pith2026250617140,
  author       = {Pith},
  title        = {Pith review of: MeDi: Metadata-Guided Diffusion Models for Mitigating Biases in Tumor Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R33YDG22}},
  note         = {Machine review of arXiv:2506.17140}
}
read the original abstract

Deep learning models have made significant advances in histological prediction tasks in recent years. However, for adaptation in clinical practice, their lack of robustness to varying conditions such as staining, scanner, hospital, and demographics is still a limiting factor: if trained on overrepresented subpopulations, models regularly struggle with less frequent patterns, leading to shortcut learning and biased predictions. Large-scale foundation models have not fully eliminated this issue. Therefore, we propose a novel approach explicitly modeling such metadata into a Metadata-guided generative Diffusion model framework (MeDi). MeDi allows for a targeted augmentation of underrepresented subpopulations with synthetic data, which balances limited training data and mitigates biases in downstream models. We experimentally show that MeDi generates high-quality histopathology images for unseen subpopulations in TCGA, boosts the overall fidelity of the generated images, and enables improvements in performance for downstream classifiers on datasets with subpopulation shifts. Our work is a proof-of-concept towards better mitigating data biases with generative models.

Figures

Figures reproduced from arXiv: 2506.17140 by the authors.

Figure 1
Figure 1. MeDi Training and Inference Framework. During training (1), the model receives real images along with their class labels and metadata. At inference (2), users can condition on arbitrary combinations of class and metadata, enabling the genera￾tion of synthetic images for underrepresented subpopulations or unseen subpopulations altogether. improved performance on the CRC-100K dataset. Aversa et al. [2] extended his￾to… view at source ↗
Figure 2
Figure 2. Dataset statistics for TCGA-UT. Left: The number of patches for the 18 largest cancer types (vertical axis) and tissue source sites (horizontal axis), capped at 1,600. Right: Race and gender distributions for the patches of selected cancer types and tissue source sites. The dataset is highly imbalanced across classes, hospitals, and demographics, with many missing or underrepresented metadata combinations. which is … view at source ↗
Figure 3
Figure 3. FID per cancer type at 800K optimization steps sampled with DDIM [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: LUSC and LUAD samples from different tissue source sites (TSS) [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Towards Robust Foundation Models for Digital Pathology

    eess.IV 2025-07 conditional novelty 6.0 of 10

    PathoROB shows that all 20 evaluated pathology foundation models encode medical center information and that lower robustness correlates with larger downstream performance drops.

Reference graph

Works this paper leans on

27 extracted references · 11 canonical work pages · cited by 1 Pith paper

  1. [1]

    Drexlin et al

    Alber, M., Tietz, S., Dippel, J., Milbich, T., Lesort, T., Korfiatis, P., Krügener, M., Cancer, B.P., Shah, N., Möllers, A., et al.: Atlas: A novel pathology foundation 10 D. Drexlin et al. model by mayo clinic, charité, and aignostics. arXiv preprint arXiv:2501.05409 (2025)

    arXiv 2025

  2. [2]

    Advances in Neural Information Processing Systems36(2024)

    Aversa, M., Nobis, G., Hägele, M., Standvoss, K., Chirica, M., Murray-Smith, R., Alaa, A.M., Ruff, L., Ivanova, D., Samek, W., et al.: Diffinfinite: Large mask-image synthesis via parallel random patch diffusion in histopathology. Advances in Neural Information Processing Systems36(2024)

    work page 2024

  3. [3]

    Nature medicine25(8), 1301–1309 (2019)

    Campanella, G., Hanna, M.G., Geneslaw, L., Miraflor, A., Werneck Krauss Silva, V., Busam, K.J., Brogi, E., Reuter, V.E., Klimstra, D.S., Fuchs, T.J.: Clinical- grade computational pathology using weakly supervised deep learning on whole slide images. Nature medicine25(8), 1301–1309 (2019)

    2019

  4. [4]

    bioRxiv pp

    Carrillo-Perez, F., Pizurica, M., Zheng, Y., Nandi, T.N., Madduri, R., Shen, J., Gevaert, O.: Rna-to-image multi-cancer synthesis using cascaded diffusion models. bioRxiv pp. 2023–01 (2023)

    work page 2023

  5. [5]

    Nature Medicine30(3), 850–862 (2024)

    Chen, R.J., Ding, T., Lu, M.Y., Williamson, D.F., Jaume, G., Song, A.H., Chen, B., Zhang, A., Shao, D., Shaban, M., et al.: Towards a general-purpose foundation model for computational pathology. Nature Medicine30(3), 850–862 (2024)

    2024

  6. [6]

    Nature medicine25(10), 1519–1525 (2019)

    Courtiol, P., Maussion, C., Moarii, M., Pronier, E., Pilcer, S., Sefta, M., Manceron, P., Toldo, S., Zaslavskiy, M., Le Stang, N., et al.: Deep learning-based classification of mesothelioma improves prediction of patient outcome. Nature medicine25(10), 1519–1525 (2019)

    work page 2019

  7. [7]

    Medical Image Analysis70, 102004 (2021)

    de Bel, T., Bokhorst, J.M., van der Laak, J., Litjens, G.: Resid- ual cyclegan for robust domain transformation of histopatho- logical tissue slides. Medical Image Analysis70, 102004 (2021). https://doi.org/https://doi.org/10.1016/j.media.2021.102004, https://www.sciencedirect.com/science/article/pii/S1361841521000505

    arXiv 2021

  8. [8]

    European journal of cancer211, 114292 (2024)

    Dernbach, G., Kazdal, D., Ruff, L., Alber, M., Romanovsky, E., Schallenberg, S., Christopoulos, P., Weis, C.A., Muley, T., Schneider, M.A., et al.: Dissecting ai- based mutation prediction in lung adenocarcinoma: A comprehensive real-world study. European journal of cancer211, 114292 (2024)

    work page 2024

Show all 27 references
  1. [9]

    Advances in neural information processing systems34, 8780–8794 (2021)

    Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. Advances in neural information processing systems34, 8780–8794 (2021)

    2021

  2. [10]

    arXiv preprint arXiv:2401.04079 (2024)

    Dippel, J., Feulner, B., Winterhoff, T., Milbich, T., Tietz, S., Schallenberg, S., Dernbach, G., Kunft, A., Heinke, S., Eich, M.L., et al.: Rudolfv: a foundation model by pathologists for pathologists. arXiv preprint arXiv:2401.04079 (2024)

    2024 arXiv

  3. [11]

    NEJM AI1(11), AIoa2400468 (2024)

    Dippel, J., Prenißl, N., Hense, J., Liznerski, P., Winterhoff, T., Schallenberg, S., Kloft,M.,Buchstab,O.,Horst,D.,Alber,M.,etal.:Ai-basedanomalydetectionfor clinical-grade histopathological diagnostics. NEJM AI1(11), AIoa2400468 (2024)

    2024

  4. [12]

    Nature Machine In- telligence2(11), 665–673 (2020)

    Geirhos, R., Jacobsen, J.H., Michaelis, C., Zemel, R., Brendel, W., Bethge, M., Wichmann, F.A.: Shortcut learning in deep neural networks. Nature Machine In- telligence2(11), 665–673 (2020)

    2020

  5. [13]

    Nature communications12(1), 4423 (2021)

    Howard, F.M., Dolezal, J., Kochanny, S., Schulte, J., Chen, H., Heij, L., Huo, D., Nanda, R., Olopade, O.I., Kather, J.N., et al.: The impact of site-specific digital histology signatures on deep learning model accuracy and bias. Nature communications12(1), 4423 (2021)

    2021

  6. [14]

    arXiv preprint arXiv:2406.16192 (2024)

    Jaume, G., Doucet, P., Song, A.H., Lu, M.Y., Almagro-Pérez, C., Wagner, S.J., Vaidya, A.J., Chen, R.J., Williamson, D.F., Kim, A., et al.: Hest-1k: A dataset for spatial transcriptomics and histology image analysis. arXiv preprint arXiv:2406.16192 (2024)

    2024 arXiv

  7. [15]

    de Jong, E.D., Marcus, E., Teuwen, J.: Current pathology foundation models are unrobust to medical center differences (2025), https://arxiv.org/abs/2501.18055 Metadata-Guided Diffusion Models for Histopathology 11

    2025 arXiv

  8. [16]

    Nature Machine Intelligence pp

    Kauffmann, J., Dippel, J., Ruff, L., Samek, W., Müller, K.R., Montavon, G.: Ex- plainable ai reveals clever hans effects in unsupervised learning models. Nature Machine Intelligence pp. 1–11 (2025)

    2025

  9. [17]

    Annual Re- view of Pathology: Mechanisms of Disease19(Volume 19, 2024), 541– 570 (2024), https://www.annualreviews.org/content/journals/10.1146/annurev- pathmechdis-051222-113147

    Klauschen, F., Dippel, J., Keyl, P., Jurmeister, P., Bockmayr, M., Mock, A., Buchstab, O., Alber, M., Ruff, L., Montavon, G., Müller, K.R.: To- ward explainable artificial intelligence for precision pathology. Annual Re- view of Pathology: Mechanisms of Disease19(Volume 19, 20...

    2024 doi

  10. [18]

    arXiv preprint arXiv:2411.05489 (2024)

    Kömen, J., Marienwald, H., Dippel, J., Hense, J.: Do histopathological foun- dation models eliminate batch effects? a comparative study. arXiv preprint arXiv:2411.05489 (2024)

    2024 arXiv

  11. [19]

    Cell Reports38(9), 110424 (2022)

    Komura, D., Kawabe, A., Fukuta, K., Sano, K., Umezaki, T., Koda, H., Suzuki, R., Tominaga, K., Ochi, M., Konishi, H., Masakado, F., Saito, N., Sato, Y., Onoyama, T., Nishida, S., Furuya, G., Katoh, H., Yamashita, H., Kakimi, K., Seto, Y., Ushiku, T., Fukayama, M., Ishikawa, S....

    2022

  12. [20]

    Nature Medicine pp

    Ktena, I., Wiles, O., Albuquerque, I., Rebuffi, S.A., Tanno, R., Roy, A.G., Azizi, S., Belgrave, D., Kohli, P., Cemgil, T., et al.: Generative models improve fairness of medical classifiers under distribution shifts. Nature Medicine pp. 1–8 (2024)

    2024

  13. [21]

    Nature communications10(1), 1096 (2019)

    Lapuschkin, S., Wäldchen, S., Binder, A., Montavon, G., Samek, W., Müller, K.R.: Unmasking clever hans predictors and assessing what machines really learn. Nature communications10(1), 1096 (2019)

    2019

  14. [22]

    In: 2009 IEEE international symposium on biomedical imaging: from nano to macro

    Macenko, M., Niethammer, M., Marron, J.S., Borland, D., Woosley, J.T., Guan, X., Schmitt, C., Thomas, N.E.: A method for normalizing histology slides for quan- titative analysis. In: 2009 IEEE international symposium on biomedical imaging: from nano to macro. pp. 1107–1110. IE...

    2009

  15. [23]

    Computers in Biology and Medicine175, 108410 (2024)

    Niehues, J.M., Müller-Franzes, G., Schirris, Y., Wagner, S.J., Jendrusch, M., Kloor, M., Pearson, A.T., Muti, H.S., Hewitt, K.J., Veldhuizen, G.P., et al.: Using histopathology latent diffusion models as privacy-preserving dataset aug- menters improves downstream classificatio...

    2024

  16. [24]

    Diagnostics14(13), 1442 (2024)

    Osorio, P., Jimenez-Perez, G., Montalt-Tordera, J., Hooge, J., Duran-Ballester, G., Singh, S., Radbruch, M., Bach, U., Schroeder, S., Siudak, K., et al.: Latent diffusion models with image-derived annotations for enhanced ai-assisted cancer diagnosis in histopathology. Diagnos...

    2024

  17. [25]

    IEEE Computer graphics and applications21(5), 34–41 (2001)

    Reinhard, E., Adhikhmin, M., Gooch, B., Shirley, P.: Color transfer between im- ages. IEEE Computer graphics and applications21(5), 34–41 (2001)

    2001

  18. [26]

    Nature Medicine30(4), 1174– 1190 (2024)

    Vaidya, A., Chen, R.J., Williamson, D.F., Song, A.H., Jaume, G., Yang, Y., Hartvigsen, T., Dyer, E.C., Lu, M.Y., Lipkova, J., et al.: Demographic bias in misdiagnosis by computational pathology models. Nature Medicine30(4), 1174– 1190 (2024)

    2024

  19. [27]

    Na- ture medicine pp

    Vorontsov, E., Bozkurt, A., Casson, A., Shaikovski, G., Zelechowski, M., Sever- son, K., Zimmermann, E., Hall, J., Tenenholtz, N., Fusi, N., et al.: A foundation model for clinical-grade computational pathology and rare cancers detection. Na- ture medicine pp. 1–12 (2024)

    2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.