Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

HASD: Hierarchical Adaption for pathology Slide-level Domain-shift

T0 review · 3 major / 5 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 The paper claims that a slide-level hierarchical adaptation framework, combining a Sinkhorn-based domain alignment solver with slide- and patch-level consistency regularizers, transfers pathology models across centers with average gains…

desk verdict A plausible slide-level DA framework whose headline gains are undermined by test-set hyperparameter selection and missing error bars. read the letter →

arxiv 2506.23673 v2 pith:77OYCTBU submitted 2025-06-30 cs.AI

Jingsong Liu , Han Li , Chen Yang , Michael Deutges , Ario Sadafi , Xin You , Katharina Breininger , Nassir Navab
show 1 more author
Peter J. Schüffler
This is my paper · ORCID
classification cs.AI
keywords domainshiftdigitalpathologywhole-slideimagesmultipleinstancelearningoptimaltransportadaptationHER2gradingsurvivalprediction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that whole-slide domain shift in pathology—where staining, scanners, and patient populations differ between centers—can be handled at the slide level rather than the patch level. The proposed HASD framework aligns feature distributions between a source and a target center with an entropic optimal-transport solver, while two regularizers stop the alignment from distorting slide morphology and from shifting attention away from diagnostic patches. Prototype selection keeps the transport computation tractable. Across eight out-of-domain setups spanning HER2 breast cancer grading and UCEC survival prediction, HASD reports average improvements of 4.1% AUROC and 3.9% C-index over the best comparison methods, with no additional pathologist annotations. If correct, this gives hospitals a practical way to reuse models across institutions.

What carries the argument

The load-bearing object is a transformation $T$ on patch feature vectors, trained by minimizing $L_{\text{total}} = L_{\text{DAS}} + \lambda_1 L_{\text{SGIR}} + \lambda_2 L_{\text{PACR}}$. $L_{\text{DAS}}$ is an entropic optimal-transport cost computed with a Sinkhorn–Knopp solver, optionally with a partial mass relaxation term; $L_{\text{SGIR}}$ preserves the Gram matrix of each slide's patch features so slide morphology is not distorted; $L_{\text{PACR}}$ penalizes changes in the aggregator's attention weights so diagnostically relevant patches keep their focus. A label-free prototype selection clusters each slide's patches and uses cluster centroids as the transport units, reducing the cost of computing the alignment.

What would settle it

Re-run the eight out-of-domain setups with all HASD hyperparameters fixed using only the source training split, then compare target performance; if the average gains over the baselines fall below 4.1% AUROC and 3.9% C-index, the reported advantage is partly due to test-set tuning. A complementary check is to apply HASD to a new center never used in the ablation study.

Watch

Extended reading notes

Core claim

The central claim is that a slide-level domain adaptation method can outperform patch-level methods on tasks where clinical decisions depend on the whole slide. HASD learns a transformation of source-domain patch features into the target domain by minimizing a transport cost computed with a Sinkhorn–Knopp solver, subject to a Gram-matrix constraint that preserves within-slide geometry and an attention-consistency constraint that keeps the attention-based multiple instance learning aggregator focused on the same diagnostic regions after transformation. In the UCEC survival task, where the two centers have different label prevalence, a partial mass relaxation is activated so that unmatched samples can be left unaligned rather than forced to match. The reported outcomes are an average 4.1% AUROC gain for HER2 grading and a 3.9% C-index gain for survival prediction across out-of-domain setups, with smaller in-domain-to-out-of-domain performance gaps than all baselines.

Load-bearing premise

The conclusion rests on the assumption that the method's settings (number of prototypes, when to activate partial mass relaxation, and the loss weights) were fixed before seeing the out-of-domain test sets, since no separate validation split is described.

Editorial extensions

If this is right

  • Across eight source-to-target setups, HASD improves average OOD AUROC by 4.1% for HER2 grading and average OOD C-index by 3.9% for UCEC survival prediction compared with the next-best baseline.
  • The in-domain to out-of-domain performance gap drops to 6.2% AUC and 6.7% C-index, roughly half the drop of the baseline methods, meaning an adapted model degrades less when deployed at an unseen center.
  • Partial DAS, with partial mass relaxation activated only when label prevalence differs, is the component that recovers survival-prediction performance in the TCGA-to-CPTAC direction.
  • Prototype selection with up to $k=10$ prototypes per slide makes the transport computation feasible for slide-sized inputs, where full patch-to-patch optimal transport would be prohibitively expensive.
  • HASD operates directly on whole slides with an attention-based aggregator, matching the clinical setup for grading and survival tasks and avoiding the need for patch-level labels in the target domain.

Reading between the lines

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

  • A stricter evaluation would freeze all HASD hyperparameters ($k$, the regularization weights, and the partial-DAS trigger) using only source-domain validation before measuring target performance, because the paper does not describe a separate validation split.
  • The same hierarchical objective could be applied to other weakly supervised slide-level tasks, such as molecular-subtype classification, where whole-slide morphology and attention patterns carry the signal.
  • Because HASD works on features from a frozen pretrained encoder, its gains may depend on the encoder; testing different foundation models would show whether the optimal prototype count changes with the feature space.
  • The partial mass relaxation could be made automatic by estimating label-prevalence ratios between centers rather than switching it on manually, which would remove one source of test-set dependence.
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

3 major / 5 minor

Summary. The paper proposes HASD, a slide-level domain adaptation framework for whole-slide pathology images. It combines a domain-level optimal-transport alignment solver (DAS) with a slide-level Gram-matrix regularization (SGIR) and a patch-level attention consistency regularization (PACR), together with a prototype selection mechanism to reduce memory cost. The method is evaluated on HER2 grading across Yale, TCGA, and TUM, and on UCEC survival prediction across TCGA and CPTAC. The authors report an average 4.1% AUROC gain in HER2 grading and a 3.9% C-index gain in survival prediction over the compared baselines, while preserving in-domain performance, and they release code.

Significance. If the reported gains are robust, HASD would be a practical contribution to slide-level domain adaptation in computational pathology: it operates at the slide level rather than the patch level, incorporates multiple regularizers to preserve structure and attention patterns, uses a label-free prototype mechanism for efficiency, and builds on a foundation-model feature extractor. The manuscript also has the strength of releasing code, which supports reproducibility. However, the current evaluation protocol leaves important questions about how the reported improvements would transfer to a genuinely new deployment site, because several key configuration choices are made using the same out-of-domain test data that is later used to compute the headline numbers.

major comments (3)
  1. [Section 3.2, Table 2] The central evaluation protocol selects hyperparameters using the OOD test setups themselves. The ablation in Table 2 chooses the prototype count k=10, the inclusion of PACR and SGIR, and the activation of partial DAS based on the OOD average scores shown in the same table (for example, k=10 is preferred over k=5 partly because it gives 59.1 vs 58.6 in Survival Prediction, and partial DAS is then added after observing a further gain to 61.4). No separate validation split drawn from the target domains is described anywhere in Section 3. Because the same setups that define the evaluation are used to select the configuration, the reported 4.1% AUROC and 3.9% C-index gains are at risk of reflecting test-set model selection rather than the method's intrinsic robustness. I ask the authors to either add a nested validation protocol with a genuinely held-out target split, or to report results for fixed default hyperparameters and show that the gains are not due to tuning on the test cohorts.
  2. [Section 2.2, Eq. (1)] The activation of partial DAS is not label-free in its current description. The text states that the partial mass relaxation term is 'only activated when label prevalence discrepancies occur across centers.' Detecting such discrepancies from the target cohort requires access to target labels (e.g., survival status in TCGA-UCEC and CPTAC-UCEC), which are exactly the labels used for evaluation in the survival task. In a real deployment scenario, target labels are unavailable at adaptation time, so the reported 3.9% survival gain cannot be expected from the algorithm as described. The authors should either present partial DAS as a semi-supervised variant that explicitly uses target label prevalence, or evaluate a fully unsupervised version that does not use any target-label information, and state clearly which configuration produces each headline number.
  3. [Section 3, Fig. 3 and Table 1] No error bars, confidence intervals, or significance tests are reported. The headline comparisons are point estimates from small cohorts (e.g., TUM has n=77 slides and Yale has n=192), so a 4.1% average AUROC difference or a 6.2% vs 10.0% gap could be within sampling variability. The paper should report repeated runs or bootstrapped confidence intervals, and ideally a statistical test or an explicit statement of effect size with uncertainty, before the claims of consistent improvement over the next-best baselines can be assessed.
minor comments (5)
  1. [Section 2.2, Eq. (1)] The displayed formula for L_DAS has mismatched parentheses and uses the symbol gamma both for the transport plan and for the entropic-regularization coefficient; please clarify the notation and write the objective without ambiguity.
  2. [Title and throughout] The manuscript inconsistently uses 'Adaption' and 'Adaptation'; the title and abstract should use one spelling consistently.
  3. [Table 2] The table header contains the typo 'Survial Prediction', and the 'DAS' row is formatted differently from the other rows; also, the reported memory usage of 52.9 GB for k=10 appears inconsistent with the statement that GPU memory limits k to 10, so please specify the hardware and memory measurement details.
  4. [Fig. 3] Several axis labels appear garbled (e.g., 'T GCA→T U M'); please correct the typesetting of the center abbreviations.
  5. [Section 3.3, Fig. 4] The t-SNE panels are small and hard to read; please enlarge them and add a clear caption explaining whether the Kaplan-Meier curve is used only for post-hoc illustration or is actually part of the decision to activate partial DAS.

Circularity Check

2 steps flagged · score 6.0 of 10

OOD gains are partly fitted to the test cohorts: Table 2 selects k and components from OOD averages, and partial DAS is triggered by target-label prevalence.

  1. fitted input called prediction [Section 2.3 and Section 3.2, Table 2 (Ablation Study for Parameter Selection)]
    "To be computationally efficient, we cluster each slide’s patches into k groups and use the clusters’ centroids as label-free prototypes (see ablation for k-selection) before applying HASD. We summarize the impact of hierarchical components, including DAS, PACR, and SGIR, and the number of prototypes on both slide-level tasks in Table 2."

    The ablation table reports OOD averages on the same target cohorts used for the headline 4.1% AUROC and 3.9% C-index numbers. The final configuration (k=10, PACR, SGIR, and partial DAS for survival) is the row with the best OOD averages, and no separate validation split is described. Thus the reported 'prediction' is partly forced by selecting model components and prototype count from the evaluation data itself.

  2. fitted input called prediction [Section 2.2, Domain-level Alignment Solver (Partial DAS)]
    "we only activate the partial mass relaxation term (i.e. set τ̸=0) when label prevalence discrepancies occur across centers. In such cases (Fig. 4 for an example), we refer to the resulting approach as Partial DAS."

    The trigger for partial DAS is the presence of label-prevalence discrepancies, which is detected from the target cohort's survival labels (the KM curves in Fig. 4a). The C-index improvement attributed to partial DAS is then measured on those same target labels. Consequently, the 3.9% survival gain is contingent on information from the test cohort's labels; in a real deployment those labels are unavailable, so the gain is not a prediction from the described algorithm alone.

full rationale

The central claim is that HASD improves OOD slide-level AUROC and C-index by 4.1% and 3.9% without extra pathologist annotations. The method itself is not definitionally equivalent to the evaluation: the DAS, SGIR, and PACR losses are stated independently of the test labels, and the baselines are external methods. However, the reported advantage is partly an artifact of model selection on the test cohorts. Section 3.2 presents Table 2 as an 'Ablation Study for Parameter Selection' whose rows are averaged over the same OOD setups used for the headline numbers; the final configuration is the row with the best averages, and Section 2.3 explicitly directs readers to this ablation for k-selection. Because no validation split is described, choosing k=10 and adding PACR, SGIR, and partial DAS from these OOD averages means the 'prediction' is partially fitted to the evaluation data. This is reinforced by the partial-DAS rule in Section 2.2, which activates the relaxation only when label-prevalence discrepancies are detected from the target cohort, information derived from the same survival labels used to compute the C-index. There is no machine-checked certificate or independent external validation that would make these choices parameter-free. I do not see a self-citation chain or uniqueness-theorem import; the circularity is specifically the test-set model selection. Score 6 reflects partial circularity: the method has independent content, but the headline gains are not fully supported as predictions because the configuration is chosen from the target test distributions.

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

The method introduces no new physical entities. It relies on several fitted hyperparameters and domain assumptions about feature transferability. The most consequential free parameters (k and tau) are selected using the test data, which inflates the circularity burden.

free parameters (4)
  • lambda1, lambda2 = not reported
    Weights for slide-level geometric invariance and patch-level attention consistency losses in L_total; no values provided in paper.
  • epsilon (entropic regularization) = 0.001
    Fixed value for Sinkhorn-Knopp solver, stated in Section 2.2.
  • tau (partial mass relaxation) = not reported; toggled on/off
    Activated only when label prevalence discrepancies are observed, a test-dependent choice (Section 2.2).
  • k (number of prototypes per slide) = 10
    Selected from ablation Table 2 based on average OOD performance; memory constrained to max 10.
assumptions (3)
  • standard math Sinkhorn-Knopp algorithm converges to the entropic OT solution for the given cost
    Invoked in Section 2.2 to compute transport costs efficiently.
  • domain assumption A single transformation T on patch features can align domain distributions while preserving slide morphology and attention patterns
    This is the modeling assumption behind the multi-loss objective; if false, the adaptation would distort diagnostic structure.
  • domain assumption UNI foundation model features are adequate for capturing slide-level diagnostic information across centers
    Patch features are fixed from UNI with no fine-tuning; the method's success depends on these features being transferable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HASD: Hierarchical Adaption for pathology Slide-level Domain-shift." pith.science (2026). https://pith.science/paper/77OYCTBU

@misc{pith2026250623673,
  author       = {Pith},
  title        = {Pith review of: HASD: Hierarchical Adaption for pathology Slide-level Domain-shift},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/77OYCTBU}},
  note         = {Machine review of arXiv:2506.23673}
}
read the original abstract

Domain shift is a critical problem for pathology AI as pathology data is heavily influenced by center-specific conditions. Current pathology domain adaptation methods focus on image patches rather than WSI, thus failing to capture global WSI features required in typical clinical scenarios. In this work, we address the challenges of slide-level domain shift by proposing a Hierarchical Adaptation framework for Slide-level Domain-shift (HASD). HASD achieves multi-scale feature consistency and computationally efficient slide-level domain adaptation through two key components: (1) a hierarchical adaptation framework that integrates a Domain-level Alignment Solver for feature alignment, a Slide-level Geometric Invariance Regularization to preserve the morphological structure, and a Patch-level Attention Consistency Regularization to maintain local critical diagnostic cues; and (2) a prototype selection mechanism that reduces computational overhead. We validate our method on two slide-level tasks across five datasets, achieving a 4.1\% AUROC improvement in a Breast Cancer HER2 Grading cohort and a 3.9\% C-index gain in a UCEC survival prediction cohort. Our method provides a practical and reliable slide-level domain adaption solution for pathology institutions, minimizing both computational and annotation costs.

Figures

Figures reproduced from arXiv: 2506.23673 by the authors.

Figure 1
Figure 1. (a)(b): Existing methods rely on patch-level labels from the source domain [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. HASD is designed for scenarios where UNI model extracts the patch [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. OOD Performances on HER2 Grading (red) and Survival Prediction (cyan). Survival Average TCGA → CPTAC TCGA → TUM Yale → TUM Yale → TCGA Grading Average TUM → TCGA TUM → Yale TCGA → Yale CPTAC → TCGA SCDA ReinhardNorm Our Method MI-SimpleShot ABMIL Her2 Grading Task Survival Prediction Task Measuring Slide-level Domain Adaption. We report the performance of all evaluated methods under various OOD setups (see [PITH_FU… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: T-SNE for the Survival task. (a) The TCGA data contain more long [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Continue with ORCID 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. From Pixels to Pathology: Restoration Diffusion for Diagnostic-Consistent Virtual IHC

    eess.IV 2025-08 reject novelty 5.0 of 10

    Star-Diff, a dual-path restoration diffusion model for virtual HER2 staining, plus the Semantic Fidelity Score metric, is demonstrated on the BCI breast cancer dataset.

Reference graph

Works this paper leans on

28 extracted references · 20 canonical work pages · cited by 1 Pith paper

  1. [21]

    Biomedical Signal Processing and Control73, 103400 (2022)

    Wang, P., Li, P., Li, Y., Xu, J., Jiang, M.: Classification of histopathological whole slide images based on multiple weighted semi-supervised domain adapta- tion. Biomedical Signal Processing and Control73, 103400 (2022)

    work page 2022

  2. [1]

    arXiv preprint arXiv:2412.04260 (2024)

    Carretero,I.,Meseguer,P.,delAmor,R.,Naranjo,V.:Enhancingwholeslideimage classification through supervised contrastive domain adaptation. arXiv preprint arXiv:2412.04260 (2024)

    arXiv 2024

  3. [2]

    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

  4. [3]

    Ad- vances in neural information processing systems26(2013)

    Cuturi, M.: Sinkhorn distances: Lightspeed computation of optimal transport. Ad- vances in neural information processing systems26(2013)

    2013

  5. [4]

    Journal of proteome research14(6), 2707–2713 (2015)

    Edwards, N.J., Oberti, M., Thangudu, R.R., Cai, S., McGarvey, P.B., Jacob, S., Madhavan, S., Ketchum, K.A.: The cptac data portal: a resource for cancer pro- teomics research. Journal of proteome research14(6), 2707–2713 (2015)

    work page 2015

  6. [5]

    Domain adaptation using optimal transport for invariant learning using histopathology datasets

    Falahkheirkhah, K., Lu, A., Alvarez-Melis, D., Huynh, G.: Domain adaptation using optimal transport for invariant learning using histopathology datasets. arXiv preprint arXiv:2303.02241 (2023)

    work page Pith review arXiv 2023

  7. [6]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Fan, J., Lv, T., Di, Y., Li, L., Pan, X.: Pathmamba: Weakly supervised state space model for multi-class segmentation of pathology images. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 500–509. Springer (2024)

    work page 2024

  8. [7]

    Modern Pathology35(1), 44–51 (2022)

    Farahmand, S., Fernandez, A.I., Ahmed, F.S., Rimm, D.L., Chuang, J.H., Reisen- bichler, E., Zarringhalam, K.: Deep learning trained on hematoxylin and eosin tumor region of interest predicts her2 status and trastuzumab treatment response in her2+ breast cancer. Modern Pathology35(1), 44–51 (2022)

    work page 2022

Show all 28 references
  1. [8]

    Journal of machine learning research17(59), 1–35 (2016)

    Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., March, M., Lempitsky, V.: Domain-adversarial training of neural networks. Journal of machine learning research17(59), 1–35 (2016)

    2016

  2. [9]

    In: BVM Workshop

    Ganz, J., Puget, C., Ammeling, J., Parlak, E., Kiupel, M., Bertram, C.A., Breininger, K., Klopfleisch, R., Aubreville, M.: Assessment of scanner domain shifts in deep multiple instance learning. In: BVM Workshop. pp. 137–142. Springer (2024)

    2024

  3. [10]

    In: International conference on machine learning

    Ilse,M.,Tomczak,J.,Welling,M.:Attention-baseddeepmultipleinstancelearning. In: International conference on machine learning. pp. 2127–2136. PMLR (2018) 10 Liu et al

    2018

  4. [11]

    arXiv preprint arXiv:2501.18055 (2025)

    de Jong, E.D., Marcus, E., Teuwen, J.: Current pathology foundation models are unrobust to medical center differences. arXiv preprint arXiv:2501.18055 (2025)

    2025 arXiv

  5. [12]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops

    Liu, S., Zhu, C., Xu, F., Jia, X., Shi, Z., Jin, M.: Bci: Breast cancer immuno- histochemical image generation through pyramid pix2pix. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops. pp. 1815–1824 (June 2022)

    2022

  6. [13]

    In: International Confer- ence on Medical Image Computing and Computer-Assisted Intervention

    Naranjo, V.: Domain adaptation for unsupervised cancer detection: An application for skin whole slides images from an interhospital dataset. In: International Confer- ence on Medical Image Computing and Computer-Assisted Intervention. Springer (2024)

    2024

  7. [14]

    Foundations and Trends®in Machine Learning11(5-6), 355–607 (2019)

    Peyré, G., Cuturi, M., et al.: Computational optimal transport: With applications to data science. Foundations and Trends®in Machine Learning11(5-6), 355–607 (2019)

    2019

  8. [15]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Poceviči¯ ut˙ e, M., Eilertsen, G., Garvin, S., Lundström, C.: Detecting domain shift in multiple instance learning for digital pathology using fréchet domain distance. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 157–167. Spr...

    2023

  9. [16]

    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

  10. [17]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Reisenbüchler, D., Luttner, L., Schaadt, N.S., Feuerhake, F., Merhof, D.: Unsu- pervised latent stain adaptation for computational pathology. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 755–765. Springer (2024)

    2024

  11. [18]

    In: International Conference on Med- ical Image Computing and Computer-Assisted Intervention

    Sharma, Y., Syed, S., Brown, D.E.: Mani: Maximizing mutual information for nu- clei cross-domain unsupervised segmentation. In: International Conference on Med- ical Image Computing and Computer-Assisted Intervention. pp. 345–355. Springer (2022)

    2022

  12. [19]

    Obstetrics & Gynecology142(4), 978–986 (2023)

    Somasegar, S., Bashi, A., Lang, S.M., Liao, C.I., Johnson, C., Darcy, K.M., Tian, C., Kapp, D.S., Chan, J.K.: Trends in uterine cancer mortality in the united states: a 50-year population-based analysis. Obstetrics & Gynecology142(4), 978–986 (2023)

    2023

  13. [20]

    Notes of Course at University of Cambridge (2018)

    Thorpe, M.: Introduction to optimal transport. Notes of Course at University of Cambridge (2018)

    2018

  14. [22]

    Wang, Z., Zhu, X., Su, L., Meng, G., Zhang, J., Li, A., Wang, M.: Instance-aware feature alignment for cross-domain cell nuclei detection in histopathology images. In: Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasb...

    2021

  15. [23]

    Nature genetics45(10), 1113–1120 (2013)

    Weinstein,J.N.,Collisson,E.A.,Mills,G.B.,Shaw,K.R.,Ozenberger,B.A.,Ellrott, K., Shmulevich, I., Sander, C., Stuart, J.M.: The cancer genome atlas pan-cancer analysis project. Nature genetics45(10), 1113–1120 (2013)

    2013

  16. [24]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Xiong, Y., Liu, J., Zaripova, K., Sharifzadeh, S., Keicher, M., Navab, N.: Prior- radgraphformer: A prior-knowledge-enhanced transformer for generating radiology graphs from x-rays. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp....

    2023

  17. [25]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Xu, Y., Chen, H.: Multimodal optimal transport-based co-attention transformer with global structure consistency for survival prediction. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 21241–21251 (2023) HASD: Hierarchical Adaption for Patholog...

    2023

  18. [26]

    arXiv preprint arXiv:2506.19742 (2025)

    Xu, Z., Li, H., Sun, D., Li, Z., Li, Y., Kong, Q., Cheng, Z., Navab, N., Zhou, S.K.: Nerf-based cbct reconstruction needs normalization and initialization. arXiv preprint arXiv:2506.19742 (2025)

    2025 arXiv

  19. [27]

    arXiv preprint arXiv:2204.08324 (2022)

    Yeaton, A., Krishnan, R.G., Mieloszyk, R., Alvarez-Melis, D., Huynh, G.: Hier- archical optimal transport for comparing histopathology datasets. arXiv preprint arXiv:2204.08324 (2022)

    2022 arXiv

  20. [28]

    In: International Conference on Medical Image Com- puting and Computer-Assisted Intervention

    Yin, C., Liu, S., Wong, V.W.S., Yuen, P.C.: Histosyn: Histomorphology-focused pathology image synthesis. In: International Conference on Medical Image Com- puting and Computer-Assisted Intervention. pp. 200–210. Springer (2024)

    2024

Pith tools

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