Abstract

Pathology foundation models (PFMs) have emerged as a core approach for learning transferable representations from whole slide images (WSIs), and they are typically benchmarked through downstream clinical endpoints. While such task level evaluations are indispensable, they offer limited insight into what the representations themselves encode, particularly whether PFM embeddings can distinguish meaningful tissue regions and capture their spatial relationships. We present SpaPath-Bench, a representation level benchmark designed to diagnose spatial representation capability in PFMs. SpaPath-Bench formulates spatial domain identification (SDI) on paired whole slide image and spatial transcriptomics (ST) data as a diagnostic task. It curates 42 public paired WSI and ST slides, enables large scale evaluation across 19 encoders and seven SDI methods, and measures partition quality using three complementary criteria: unsupervised spatial coherence, transcriptomics referenced agreement, and expe

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2389_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://bokai-zhao.github.io/SpaPath-benchboard/

Link to the Dataset(s)

N/A

BibTex

@InProceedings{ZhaBok_Benchmarking_MICCAI2026,
        author = { Zhao, Bokai AND Zhang, Yiyang AND Zhu, Yuanchi AND Chao, Hanqing AND Bai, Long AND Ma, Tai AND Xu, Minfeng AND Song, Ming AND Jiang, Tianzi},
        title = { { Benchmarking Pathology Foundation Models for Spatial Domain Understanding } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16895},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper introduces a benchmark studying how embeddings from tile-level pathology foundation models can be aggregated to distinguish tissue regions. More specifically, they treat spatial domain identification (SDI) as an evaluation task to compare different foundation models, SDI and clustering methods. The benchmark ranks foundation models, SDI and clustering methods based on a large set of metrics and using both expert annotation and spatial transcriptomic expressions as ground-truth annotations.

  • 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.
    • S1.Studying how to build spatial representations from pathology foundation model embeddings is valuable for the community.
    • S2.Various foundation models, SDI and clustering methods are covered.
    • S3.The paper draws conclusion from a various set of experiments involving a large number of metrics and varying ground-truth sources (expert annotations, ST expressions).
  • 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.
    • W1.The paper claims to evaluate “whether PFM embeddings can distinguish meaningful tissue regions and capture their spatial relationships” (quote from abstract). However, the considered foundation models are tile-level models that extract embeddings only for a small sub-region of a slide. It is thus misleading to present this benchmark as a way to valuate the spatial understanding of models. Indeed, one could argue that the spatial understanding is performed when aggregating tile-level embeddings into slide-level regions. It is thus rather the SDI methods that perform such “spatial domain understanding”.

    • W2.the PCA non-spatial baseline is very briefly presented. It is not clear to me how it is exactly implemented by the authors.

    • W3.Metrics are also very briefly introduced and there are a lot: CHAOS, PAS, ASW, ARI, NMI, HOM, COM. Presenting such metrics in more details could make the paper self-contained and make it easier to appreciate the reported results.

    • W4.Results are averaged only among 42 WSIs which is a quite low sample size. However, I acknowledge the lack of large publicly-available WSI-ST paired datasets.

  • 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 has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I believe the introduced benchmark is valuable for the community. However, as noted in the “Wekanesses” section, I have concerns about the paper claiming to evaluating the quality of spatial representations from foundation models. Also, I find the paper lacks details about certain baselines and metrics, and authors should acknoledge more explicitly the small sample size of the considered dataset.

  • Reviewer confidence

    Somewhat confident (2)

  • [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 #2

  • Please describe the contribution of the paper

    This paper proposes a new benchmark for the histopathology domain to test the spatial representation capabilities of pathology foundation models (PFMs). While many benchmarks now exist, such as HEST, THUNDER, PathBench, and even tile-level segmentation-based ones, there is a lack of slide-level representation capability benchmarks for dense tasks (in contrast to aggregation-based tasks in PathBench). The authors use complementary spatial transcriptomics data to formulate spatial domain identification (SDI) as a diagnostic task. The paper covers seven SDI metrics across three complementary regimes benchmarked for 19 PFMs.

  • 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) The quantification of spatial domain understanding via three complementary regimes with seven defined metrics is exhaustive and relevant for measuring the spatial representation capabilities across expert knowledge, concordance with molecular profiles, and intrinsic spatial coherence.

    2) The benchmark includes a total of 19 PFMs, covering pure vision, vision-language, natural image, and even vision-omics models, making the experimental setup in this study stronger. Overall, this benchmark could greatly benefit the histopathology community.

  • 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) The proposed benchmark, while meant for testing spatial representation capabilities, solely relies on spatial transcriptomics data. There exist multiple patch-level segmentation datasets for whole WSIs [1-2]; therefore, those could have been added as well to test purely histology-only, morphology-based downstream tasks.

    2) In Section 4.3, the authors discuss the limitation of the local H&E approach alone in PFMs for suboptimally encoding high-level neuroanatomical context. However, there are now numerous transformer-based slide-level encoders (acting as contextualizers), which can enhance tile-level embeddings with the context of the WSIs. Benchmarking at least a couple of slide-encoder outputs would make this benchmark even richer. Some examples that the authors can include are TICON, Gigapath-SE (slide-encoder), TITAN, etc.

  • 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.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/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.

    (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 benchmark proposed is quite exhaustive in terms of metrics, tasks, and models benchmarked. The insights are relavant and can guide new pathology foundation models development.

  • Reviewer confidence

    Very confident (4)

  • [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 mainly contributes a benchmark for testing pathology foundation models on spatial transcriptomics slides with paired H&E images. The proposed benchmark and evaluation:

    • Collects 42 paired H&E–ST slides
    • Includes a small expert-labeled set and a larger unlabeled set
    • Defines a common pipeline for patch extraction, embedding, spatial modeling, clustering, and evaluation
    • Compares many pathology and vision-language foundation models under the same setup

    The paper is useful because it gives a more standardized way to compare models on this task.

  • 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 benchmark setup is careful and organized.
    • Compared many models under the same pipeline, which makes the comparison more fair.
    • Using both expert-labeled slides and transcriptomics-based proxy evaluation is a practical design.
    • The paper may be valuable as a resource for the community.
  • 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.The expert-labeled evaluation is limited: Only 12 slides have expert annotations, and all come from the same DLPFC setting. As a result, the paper cannot strongly claim that the findings generalize broadly across tissues, disease settings, or pathology tasks.

    2.Most of the evaluation depends on a proxy rather than true ground truth: For the 30 unlabeled slides, the reference is based on HVG-driven transcriptomic clustering. This is a reasonable proxy, but it still reflects agreement with one transcriptomics-based pipeline rather than true biological, anatomical, or clinical structure.

    3.Some of the main findings are not very surprising: It is not unexpected that a large pathology-specific vision model performs strongly on this task, or that a vision-language model aligns better with expert-defined, human-interpretable regions. So while the benchmark is useful, the main conclusions seem more confirmatory than novel.

    4.The compared models differ along several confounded factors: These include model size, training data scale, pretraining objective, and domain specificity. Because these factors vary at the same time, it is difficult to draw strong conclusions about why one pretraining strategy works better than another.

    5.The results also depend on downstream design choices: Final performance is influenced not only by the embeddings, but also by the spatial modeling method, clustering algorithm, and the strategy used to select the number of clusters. Therefore, the paper should be careful not to attribute all performance differences solely to the foundation models.

  • 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.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • 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/2026/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 — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The benchmark is useful and the pipeline is fairly solid, but the evidence is still limited. Only a small subset has expert labels, most results depend on proxy transcriptomic partitions, and the main conclusions are not very surprising. So the paper feels more like a solid benchmark/resource paper than a paper with strong new insights.

  • 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



Author Feedback

We sincerely thank all reviewers for their constructive and encouraging comments. We are grateful that the reviewers recognized the value of our benchmark and its systematic evaluation of pathology foundation models, spatial representation modules, clustering strategies, and complementary evaluation metrics.

We also appreciate the insightful suggestions regarding the interpretation of spatial understanding in tile-level PFMs, the influence of downstream SDI and clustering choices, the use of transcriptomic proxy references, the limited number of expert-annotated slides, and the potential inclusion of histology-only dense annotations and slide-level contextualized encoders. These are important directions that will help further improve and extend the benchmark in future work.

We thank the reviewers again for their careful reading and valuable feedback. We will take these suggestions into account in our future extensions of this work.




Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    Most reviewers regarded the paper as strong and clearly recommended acceptance.



back to top