List of Papers Browse by Subject Areas Author List
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.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1963_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/TumVink/HASD
Link to the Dataset(s)
HER2 dataset: https://www.cancerimagingarchive.net/collection/her2-tumor-rois/
UCEC-TCGA dataset: https://www.cancer.gov/ccg/research/genome-sequencing/tcga
UCEC-CPTAC dataset: https://gdc.cancer.gov/about-gdc/contributed-genomic-data-cancer-research/clinical-proteomic-tumor-analysis-consortium-cptac
BibTex
@InProceedings{LiuJin_HASD_MICCAI2025,
author = { Liu, Jingsong and Li, Han and Yang, Chen and Deutges, Michael and Sadafi, Ario and You, Xin and Breininger, Katharina and Navab, Nassir and Schüffler, Peter J.},
title = { { HASD: Hierarchical Adaption for Pathology Slide-Level Domain-Shift } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
year = {2025},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15965},
month = {September},
page = {337 -- 347}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes HASD (Hierarchical Adaptation framework for Slide-level Domain-shift) to address domain shift at the WSI level in computational pathology. The main contribution lies in the design of a multi-scale domain adaptation framework that effectively aligns features across domains while preserving diagnostic cues. HASD introduces a computationally efficient prototype selection mechanism to reduce computational overhead.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
(1) Hierarchical adaptation formulation. The proposed HASD framework 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 (2) Prototype selection for computational efficiency to reduce the computational overhead. (3) Extensive empirical evaluation. The method is evaluated on two tasks: HER2 grading in breast cancer and survival prediction in uterine cancer across five datasets. HASD achieves notable performance improvements (e.g., +4.1% AUROC, +3.9% C-index), demonstrating its effectiveness.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
(1) Limited methodological novelty. While the hierarchical adaptation approach is well-structured, similar ideas have been explored in related domains. For instance, optimal transport [1] and morphological prototyping for survival analysis [2] share conceptual similarities with the proposed modules. The manuscript would benefit from a clearer differentiation from these prior works. (2) Unclear slide level representation via prototypes: The authors perform prototype selection prior to applying HASD, the process make HASD still appear to patch level adaption but not slide level. The manuscript does not clearly explain how the selected prototypes effectively capture or represent the holistic slide-level information. (3) Insufficient number of prototypes: The number of prototypes is fixed at 10, which raises concerns about whether this is sufficient to capture the diverse morphological variations present in gigapixel WSIs. (4) Lack of computational efficiency evidence: The authors claim that the prototype selection mechanism improves computational efficiency. However, the manuscript does not provide quantitative evidence (e.g., runtime comparisons, memory usage) to support this claim. (5) Limited baseline comparisons: The empirical evaluation could be significantly strengthened by including more recent and relevant state-of-the-art methods. The current set of baselines may not fully reflect the progress in domain adaptation for computational pathology, which may limit the perceived generalizability and impact of the proposed approach. [1] Xu, Yingxue, and Hao Chen. “Multimodal optimal transport-based co-attention transformer with global structure consistency for survival prediction.” Proceedings of the IEEE/CVF international conference on computer vision. 2023. [2] Song, Andrew H., et al. “Morphological prototyping for unsupervised slide representation learning in computational pathology.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — could be rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The overall rating is primarily influenced by the limited novelty of the proposed method and the lack of sufficient experimental justification for key design choices (e.g., prototype selection strategy, computational efficiency). While the framework is well-motivated and the experiments are thorough, the current version of the paper does not convincingly demonstrate a substantial advancement over existing approaches. I am open to increasing my rating if my concerns are well-addressed.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
Thank you for the detailed responses. My major concerns are addressed, I tend to accept this manucript.
Review #2
- Please describe the contribution of the paper
This paper proposes HASD to address the slide-level domain transfer problem. Results on two public datasets demonstrate that HASD helps improve the generalization ability of external tests.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper has goodmethodological advantages. A large-scale of experiments are used to verify the idea. The performance of the model is outstanding.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
I have a few concerns:
Does the number of source and target patches required by this domain-adaptive method affect the training of HASD? Some clarification on this point would be helpful.
Several recent slide-level foundation models[1][2] also aim to address the domain adaptation problem. How does HASD compare to these models? Does it offer any clear advantages?
Few-shot learning is an effective approach for domain adaptation by fine-tuning a small number of samples. I believe that including a comparison between HASD and few-shot fine-tuning methods would strengthen the manuscript by highlighting the zero-shot capabilities of HASD.
I could not find the evaluation metrics for the first dataset in the manuscript. I recommend the authors provide a more detailed description of the quantitative evaluation metrics in the implementation section.
[1] Xu, Hanwen, et al. “A whole-slide foundation model for digital pathology from real-world data.” Nature 630.8015 (2024): 181-188. [2] Wang, Xiyue, et al. “A pathology foundation model for cancer diagnosis and prognosis prediction.” Nature 634.8035 (2024): 970-978.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — could be accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Please see the weakness part. The method proposed in this paper actually has great research potential. I look forward to the author’s reply.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
This paper targets the issue of slide-level domain shift in computational pathology, which current domain adaptation (DA) methods fail to address effectively due to their patch-level focus. The authors propose HASD, a Hierarchical Adaptation framework for slide-level domain shift. HASD combines three levels of alignment: domain-level (via a novel partial domain alignment solver), slide-level (via geometric invariance), and patch-level (via attention consistency regularization) alongside a prototype selection mechanism for computational efficiency.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The paper introduces a well-motivated, novel framework (HASD) specifically tailored for slide-level domain adaptation, which is still a bit underexplored in computational pathology (mostly due to available public dataset composition).
- The hierarchical adaptation design spans domain-, slide-, and patch-levels, allowing preservation of both global morphology and local diagnostic cues.
- The use of a Partial DAS to account for label prevalence discrepancies is a practical and thoughtful enhancement.
- Prototype selection reduces computational load while preserving essential structure, making the approach scalable and realistic for clinical WSI data.
- Experimental results are extensive and convincing, with evaluations on two clinically meaningful tasks and across five datasets, showing improvements over multiple baselines.
- Especially including a thorough ID vs. OOD analysis (Table 1), clearly demonstrates generalization performance and robustness.
- Ablation studies are well-conducted, showing the additive effect of each component and justifying design decisions (e.g., prototype count).
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- The methodology, particularly the hierarchical loss formulation and optimization, is quite complex and could benefit from more intuitive explanation or a visual schematic summarizing the full objective.
- The partial DAS relies on prior knowledge of label prevalence differences across centers, which may not always be known or easy to estimate in practice.
- No analysis is given on the computational runtime or memory usage (e.g., how prototype selection actually affects training time).
- Evaluation is limited to two tasks; while results are strong, additional evidence from other slide-level endpoints (e.g. tumor subtyping) would strenghten generalizability claims.
- There is no discussion or visualization of failure cases or limitations of the approach (e.g. what if morphological structures differ significantly between domains?).
- Code only available upon acceptance.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
The paper is well-motivated and clearly fills a gap in slide-level domain adaptation. However, several areas would benefit from clarification or expansion:
- Think about providing a visual overview of the loss components (DAS, SGIR, PACR) and how they interact during optimization.
- A quantitative report of runtime and memory impact (with/without prototype selection) would be helpful, especially for users deploying on clinical-scale WSIs.
- It would be interesting to see whether HASD performs well when domain shifts include substantial anatomical differences, not just imaging or staining variation.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses an important and underexplored challenge in computational pathology: domain shift at the slide level. The proposed method is, in my opinion, technically sound, thoughtfully constructed, and empirically validated across multiple realistic clinical settings. While the novelty lies more in integration than in individual components, the framework is impactful and highly relevant to the MICCAI community. Minor limitations in clarity and practical usability should be addressed during revision.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
Still satisfied with the work. No negative points from my side.
Author Feedback
We are excited reviewers find our work well-motivated (R1, R3, R4), highly relevant for MICCAI (R1), methodologically novel and sound (R1, R3), and experiments are thorough (R1, R3, R4), With their critical feedback, we could improve our manuscript further.
How does ProtoType (PT) number affect computational cost? (R1, R4) We report GPU usage with increasing PT number k per slide. For k = [1, 5, 10, 13, 15], the memory consumption is [0.44, 9.29, 33.89, 62.03, ~85] GB. These results are consistent with the optimal transport (OT) ‘s quadratic complexity (O(n²), Sec. 2.3). Compared to using all patches directly, our PT selection strategy significantly reduces computational cost.
How well do PTs represent the slide? (R4) To measure the fidelity of our k-means selected PTs, we compare them to random sampling via Wasserstein distance. Lower distance indicates better approximation of the full patch distribution. For k = [1, 5, 10, 13, 15], random sampling (mean of 3 runs):[19.4, 19.1, 18.9, 18.8, 18.8], ours:[8.6, 5.9, 4.9, 4.5, 4.4]. Our strategy achieves significantly lower distances and effectively captures the whole slide information.
Are 10 PTs enough? (R4): K=10 offers a strong balance between computational efficiency and representation quality. Increasing K (e.g., from 10 to 15) yields only marginal gains (see above). This choice is also biologically grounded: pathology slides typically contain <10 main tissue types (i.e., PTs). Breast cancer slides, for example, often have 7 [Mehta et al., WACV 2018].
Does total patch number affect training? (R3): HASD is unaffected by the total number of patches in slides. The PTs are selected from all patches in a slide with a fix number (e.g., 10) to train the domain alignment-layer.
Is HASD applied for slide-level or patch-level alignment? (R4) HASD performs a hierarchical slide-level domain alignment. Only PTs are used during training for efficiency, but the learned transformation is applied to all patches during inference.
How does HASD differ from related works? (R4): Our slide-level domain alignment method is fundamentally different from MOTCat [Xu et al., ICCV 2023] and PANTHER [Lu et al., CVPR 2024] in terms of objectives, challenges, and implementation. MOTCat focuses on matching the WSIs and genomics simply using OT, but it does not perform domain alignment and shares no conceptual overlap beyond the use of OT. PANTHER, use EM-based approach to find PT, while success, it is not proper for our task as it is extremely computationally intensive. In contrast, our method adopts efficient K-means–based PT selection to construct src and tgt domains, focusing on hierarchical cross domain alignment while preserving both slide structure and patch-level consistency, which is not addressed in prior works. We will clarify these distinctions in the revised manuscript.
More SOTA methods comparison? (R3, R4) We included the SOTA slide-level foundation model CHIEF [Wang et al., Nature 2024], but it shows limited performance under domain shift. On HER2 Detection (Task 1, 2 OOD setups), HASD achieves an AUC of 0.787, clearly outperforming Pretrained CHIEF (0.574), Finetuned CHIEF (0.736).
How is HASD compared to few-shot learning? (R3) We compared zero-shot HASD with few-shot ABMIL with shots=[0,1,3,5] for each class. Under the same OOD setup as above, ABMIL yields mean AUCs of [74.8, 75.3, 76.7, 79.0], while HASD achieves 78.7.
Does partial DAS require label prevalence knowledge? (R1) HASD is not sensitive to the estimated label differences, if the relation value is 0. We experimented with \tau = [0, 0.1, 0.3] on survival prediction task, and the C-Index scores are [60.1, 61.4, 61.2].
Minor comments:
We will add a figure to explain the loss components (R1), clearly state that the evaluation metric for HER2 scoring is AUC (R3), add additional slide-level tasks and discussion of failure case (R1).
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.
N/A
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
This paper receives an initial review of 1A (R1), 1WA (R3), and 1WR (R4). After rebuttal, R4 changes to A based on satisfactory responses to major concerns. The main criticisms included limited methodological novelty, unclear slide-level representation via prototypes, lack of computational efficiency evidence, and missing comparisons with recent foundation models. The authors successfully addressed these concerns in their rebuttal by providing computational cost analysis, demonstrating prototype selection effectiveness, including comparisons with SOTA foundation models like CHIEF, and clarifying methodological differences from related works, leading to the paper’s acceptance.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A