The reviewed record of science sign in
Pith

arxiv: 2605.24253 · v3 · pith:ASSUKOAR · submitted 2026-05-22 · cs.CV · cs.AI· cs.IR

CRISP -- Clustering-Based Redundancy-Reduced Instance Sampling for Pathology Case Representation and Retrieval

Zahra Rahimi Afzal , Wataru Uegami , Saghir Alfasly , Wenchao Han , Saba Yasir , Judy C. Boughey , Matthew P. Goetz , Krishna R. Kalari
show 1 more author
H.R. Tizhoosh
This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-06-30 15:24 UTCgrok-4.3pith:ASSUKOARrecord.jsonopen to challenge →

classification cs.CV cs.AIcs.IR
keywords digital pathologywhole slide imagesclustering samplingcase retrievalbreast cancerredundancy reductionmulti-WSI analysis
0
0 comments X

The pith

CRISP uses clustering to build case representations from multiple pathology slides that match or exceed single-slide selection for retrieval.

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

Digital pathology cases typically involve multiple whole-slide images showing different parts of a tumor, yet analysis often depends on just one slide chosen by a pathologist. The paper proposes CRISP, which first cuts down on similar patches inside each slide and then clusters patches across all slides to pick a small representative group. This group captures the case's diversity and acts as the basis for finding similar cases. Tests on two breast cancer datasets show CRISP performs at least as well as the usual method of combining AI models with pathologist choices. By handling all slides automatically, the method can use information that single-slide approaches miss.

Core claim

CRISP is an unsupervised two-stage framework that reduces redundancy within individual whole-slide images and then performs clustering-based sampling to select a compact set of patches representing the full case, which serves as a retrieval index and achieves performance matching or surpassing combined model and pathologist slide selection on breast cancer datasets.

What carries the argument

Clustering-Based Redundancy-Reduced Instance Sampling (CRISP): a two-stage process of intra-WSI redundancy reduction followed by inter-WSI clustering to select representative patches for case-level retrieval.

If this is right

  • Case-level analysis becomes possible without discarding information from unselected slides.
  • Retrieval accuracy can be maintained or improved while processing fewer patches overall.
  • Subjective elements in slide selection for case representation are removed.
  • Scalable handling of multi-WSI cases in large pathology archives is enabled.

Where Pith is reading between the lines

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

  • Applying similar redundancy reduction and clustering to other multi-image medical modalities could yield comparable efficiency gains.
  • Combining CRISP with additional clinical metadata might enhance retrieval relevance beyond morphology alone.
  • Testing on datasets with known diagnostic pitfalls would reveal if the sampling preserves rare but critical features.

Load-bearing premise

Clustering will select patches that preserve the clinically important morphological variations across the case's slides.

What would settle it

Finding a set of cases where retrieval using CRISP patches returns lower similarity scores or incorrect matches compared to using pathologist-selected slides, specifically because key diagnostic patches were not sampled.

Figures

Figures reproduced from arXiv: 2605.24253 by H.R. Tizhoosh, Judy C. Boughey, Krishna R. Kalari, Matthew P. Goetz, Saba Yasir, Saghir Alfasly, Wataru Uegami, Wenchao Han, Zahra Rahimi Afzal.

Figure 1
Figure 1. Figure 1: CRISP workflow for case-level histopathology representation from multi-slide whole-slide images: (A) Number of whole-slide images (WSIs) per case in the diagnostic study cohort. (B) Number of WSIs per diagnostic subtype, showing class imbalance across categories. (C) Number of whole-slide images (WSIs) per triple-negative breast cancer patient in the treatment study cohort. (D) Number of patients for posit… view at source ↗
Figure 2
Figure 2. Figure 2: Results for breast morphological subtyping (50 patients): Yottixel, SPLICE, and two WSI foundation models TITAN and PRISM were applied on a single WSI selected by a pathologist as baselines; MOOZY and ABMIL are included as patient-level representation methods that aggregate all case WSIs into a single embedding without explicit patch diversity; compared to CRISP in different settings to process all case WS… view at source ↗
Figure 3
Figure 3. Figure 3: Results for prediction of treatment response for triple negative breast cancer (209 patients): Yottixel, SPLICE, and two WSI foundation models TITAN and PRISM were applied on a single WSI selected by a pathologist as baselines; MOOZY and ABMIL are included as patient-level representation methods that aggregate all case WSIs into a single embedding without explicit patch diversity; compared to CRISP in diff… view at source ↗
Figure 4
Figure 4. Figure 4: Top: Whole-slide image (WSI) for patient 1 from 50-patient dataset with the subtype Invasive carcinoma, NST. Bottom: Corresponding patch extracted from the rectangular region selected in the top image at 20× magnification with a size of 1000×1000 pixels. 13/16 [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Top: Whole-slide image (WSI) for patient 33 from 50-patient dataset with the subtype Ductal carcinoma in situ. Bottom: Corresponding patch extracted from the rectangular region selected in the top image at 20× magnification with a size of 1000×1000 pixels. 14/16 [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Top: Whole-slide image (WSI) for patient 1 from TNBC dataset with non-pCR. Bottom: Corresponding patch extracted from the rectangular region selected in the top image at 20× magnification with a size of 1000×1000 pixels. 15/16 [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Top: Whole-slide image (WSI) for patient 21 from TNBC dataset with with pCR. Bottom: Corresponding patch extracted from the rectangular region selected in the top image at 20× magnification with a size of 1000×1000 pixels. 16/16 [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
read the original abstract

Digital pathology archives increasingly contain multiple whole-slide images (WSIs) per case, capturing spatially distinct tumor regions and reflecting intrinsic morphological heterogeneity. However, most existing approaches rely on a single pathologist-selected slide, thereby discarding potentially informative evidence distributed across the remaining WSIs. To date, no autonomous framework has been proposed for comprehensive multi-WSI case processing. Here, we present an unsupervised framework for case-level analysis that integrates information from all available slides within a case. Rather than relying on a single designated slide, the proposed approach constructs case-level representations by selectively distilling informative patches across WSIs. We introduce Clustering-Based Redundancy-Reduced Instance Sampling for Pathology (CRISP), a two-stage framework that first reduces redundancy within individual WSIs and subsequently applies clustering-based sampling to select a compact yet representative set of patches for the entire case. The resulting patch set captures case-level heterogeneity while avoiding exhaustive processing of gigapixel images, and directly serves as a retrieval index. Using two Mayo Clinic breast cancer datasets for diagnosis and treatment planning, we demonstrate that CRISP consistently matches or surpasses the current standard practice of combined model and pathologist slide selection for patient/case search and retrieval. By automating case-level processing and eliminating subjective WSI selection, CRISP potentially enables the exploitation of clinically relevant information distributed across multiple WSIs that is currently overlooked.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces CRISP, an unsupervised two-stage framework for case-level analysis in digital pathology. It first performs redundancy reduction within each WSI and then applies clustering-based sampling across WSIs to select a compact, representative set of patches. The central claim is that the resulting case representations match or exceed the performance of the current standard (combined model and pathologist slide selection) for patient/case retrieval, as demonstrated on two Mayo Clinic breast cancer datasets for diagnosis and treatment planning.

Significance. If the empirical results hold, the work would provide a practical unsupervised method to exploit morphological heterogeneity distributed across multiple WSIs per case, removing reliance on subjective single-slide selection. The two-stage redundancy-reduction-plus-clustering design is a clear strength, as is the direct applicability to retrieval indexing without requiring exhaustive gigapixel processing.

major comments (2)
  1. [Abstract] Abstract: the central empirical claim that CRISP 'consistently matches or surpasses' standard practice is load-bearing for the paper, yet the abstract supplies no quantitative metrics, baselines, statistical tests, or method details (feature extractor, clustering criterion, or number of selected patches). The results section must contain these to allow verification of the performance comparison.
  2. [Methods] Methods (clustering stage): the two-stage process can discard minority clusters; if the embedding space does not separate clinically relevant heterogeneity, small but diagnostically critical morphological variants may be omitted. The manuscript must provide either expert alignment checks on selected patches or an ablation on cluster-size thresholds to substantiate that the sampling preserves information essential for diagnosis and treatment planning.
minor comments (1)
  1. [Abstract] Abstract: the phrase 'combined model and pathologist slide selection' is used without a precise definition of the baseline procedure; a short clarification would aid readers.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central empirical claim that CRISP 'consistently matches or surpasses' standard practice is load-bearing for the paper, yet the abstract supplies no quantitative metrics, baselines, statistical tests, or method details (feature extractor, clustering criterion, or number of selected patches). The results section must contain these to allow verification of the performance comparison.

    Authors: The results section provides the quantitative metrics, baselines (combined model and pathologist slide selection), statistical tests, feature extractor details, clustering criterion, and number of selected patches on both Mayo Clinic datasets, enabling verification of the claims. The abstract is a concise summary and does not include these specifics. revision: no

  2. Referee: [Methods] Methods (clustering stage): the two-stage process can discard minority clusters; if the embedding space does not separate clinically relevant heterogeneity, small but diagnostically critical morphological variants may be omitted. The manuscript must provide either expert alignment checks on selected patches or an ablation on cluster-size thresholds to substantiate that the sampling preserves information essential for diagnosis and treatment planning.

    Authors: We agree this is a valid concern for the unsupervised clustering stage. Expert alignment checks are outside the scope of the unsupervised framework. We will add an ablation on cluster-size thresholds to the revised manuscript to demonstrate preservation of information relevant to the retrieval tasks. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical clustering method validated externally

full rationale

The paper introduces an unsupervised two-stage sampling framework (per-WSI redundancy reduction followed by case-level clustering) and evaluates it through direct comparison of retrieval performance against pathologist-plus-model baselines on two independent Mayo Clinic breast cancer datasets. No equations, fitted parameters, self-referential derivations, or load-bearing self-citations appear in the provided text. The central claim is supported by experimental results on held-out data rather than by construction from the method's own inputs or prior author work.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities are described.

pith-pipeline@v0.9.1-grok · 5821 in / 1028 out tokens · 39570 ms · 2026-06-30T15:24:29.657048+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages · 1 internal anchor

  1. [1]

    & Lee, G

    Madabhushi, A. & Lee, G. Image analysis and machine learning in digital pathology: Challenges and opportunities.Med. Image Analysis33, 170–175, DOI: 10.1016/j.media.2016.06.037 (2016)

    work page doi:10.1016/j.media.2016.06.037 2016

  2. [2]

    Image Analysis65, 101757, DOI: 10.1016/j.media.2020.101757 (2020)

    Kalra, S.et al.Yottixel – an image search engine for large archives of histopathology whole slide images.Med. Image Analysis65, 101757, DOI: 10.1016/j.media.2020.101757 (2020). 3.Tizhoosh, H. R. & Pantanowitz, L. On image search in histopathology.J. Pathol. Informatics15, 100375 (2024)

    work page doi:10.1016/j.media.2020.101757 2020

  3. [3]

    R.et al.Searching images for consensus: can ai remove observer variability in pathology?The Am

    Tizhoosh, H. R.et al.Searching images for consensus: can ai remove observer variability in pathology?The Am. journal pathology191, 1702–1708 (2021)

    work page 2021

  4. [4]

    Lahr, I.et al.Analysis and validation of image search engines in histopathology.IEEE Rev. Biomed. Eng.18, 350–367 (2024)

    work page 2024

  5. [5]

    & Rubin, D

    Barker, J., Hoogi, A., Depeursinge, A. & Rubin, D. L. Automated classification of brain tumor type in whole-slide digital pathology images using local representative tiles.Med. Image Analysis30, 60–71, DOI: 10.1016/j.media.2015.12.002 (2016)

    work page doi:10.1016/j.media.2015.12.002 2015

  6. [6]

    & Zhang, S

    Zhang, X., Liu, W., Dundar, M., Badve, S. & Zhang, S. Towards large-scale histopathological image analysis: Hashing- based image retrieval.IEEE Transactions on Med. Imaging34, 496–506, DOI: 10.1109/TMI.2014.2363661 (2015)

    work page doi:10.1109/tmi.2014.2363661 2014

  7. [7]

    Medicine2, 56, DOI: 10.1038/ s41746-019-0131-z (2019)

    Hegde, N.et al.Similar image search for histopathology: Smily.npj Digit. Medicine2, 56, DOI: 10.1038/ s41746-019-0131-z (2019). 9.Shao, D.et al.Do multiple instance learning models transfer? InInternational conference on machine learning(2025)

    work page 2019

  8. [8]

    Y .et al.Data-efficient and weakly supervised computational pathology on whole-slide images.Nat

    Lu, M. Y .et al.Data-efficient and weakly supervised computational pathology on whole-slide images.Nat. Biomed. Eng. 5, 555–570, DOI: 10.1038/s41551-020-00682-w (2021)

    work page doi:10.1038/s41551-020-00682-w 2021

  9. [9]

    In Advances in Neural Information Processing Systems, vol

    Shao, Z.et al.Transmil: Transformer based correlated multiple instance learning for whole slide image classification. In Advances in Neural Information Processing Systems, vol. 34, 2136–2147 (2021)

    work page 2021

  10. [10]

    Medicine25, 1301–1309, DOI: 10.1038/s41591-019-0508-1 (2019)

    Campanella, G.et al.Clinical-grade computational pathology using weakly supervised deep learning on whole slide images.Nat. Medicine25, 1301–1309, DOI: 10.1038/s41591-019-0508-1 (2019)

    work page doi:10.1038/s41591-019-0508-1 2019

  11. [11]

    Towards a General-Purpose Foundation Model for Computational Pathology,

    Chen, R. J.et al.Towards a general-purpose foundation model for computational pathology.Nat. Medicine30, 850–862, DOI: 10.1038/s41591-024-02857-3 (2024)

    work page doi:10.1038/s41591-024-02857-3 2024

  12. [12]

    A visual -language Foundation Model for Computational Pathology,

    Lu, M. Y .et al.A visual-language foundation model for computational pathology.Nat. Medicine30, 863–874, DOI: 10.1038/s41591-024-02856-4 (2024)

    work page doi:10.1038/s41591-024-02856-4 2024

  13. [13]

    Medicine30, 2924–2935, DOI: 10.1038/s41591-024-03141-0 (2024)

    V orontsov, E.et al.A foundation model for clinical-grade computational pathology and rare cancers detection.Nat. Medicine30, 2924–2935, DOI: 10.1038/s41591-024-03141-0 (2024). 11/16

    work page doi:10.1038/s41591-024-03141-0 2024

  14. [14]

    Xu, H.et al.A whole-slide foundation model for digital pathology from real-world data.Nature630, 181–188, DOI: 10.1038/s41586-024-07441-w (2024)

    work page doi:10.1038/s41586-024-07441-w 2024

  15. [15]

    Wang, X.et al.A pathology foundation model for cancer diagnosis and prognosis prediction.Nature634, 970–978, DOI: 10.1038/s41586-024-07894-z (2024)

    work page doi:10.1038/s41586-024-07894-z 2024

  16. [16]

    Virchow2: Scaling self- supervised mixed magnification models in pathology

    Zimmermann, E.et al.Virchow2: Scaling self-supervised mixed magnification models in pathology.arXiv preprint arXiv:2408.00738(2024). 2408.00738

    work page arXiv 2024

  17. [17]

    MedicineDOI: 10.1038/ s41591-025-03982-3 (2025)

    Ding, T.et al.A multimodal whole-slide foundation model for pathology.Nat. MedicineDOI: 10.1038/ s41591-025-03982-3 (2025)

    work page 2025

  18. [18]

    MOOZY: A Patient-First Foundation Model for Computational Pathology

    Kotp, Y ., Trinh, V . Q.-H., Pal, C. & Hosseini, M. S. Moozy: A patient-first foundation model for computational pathology (2026). 2603.27048

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  19. [19]

    2404.17704

    Alsaafin, A.et al.Splice – streamlining digital pathology image processing.arXiv preprint arXiv:2404.17704(2024). 2404.17704

    work page arXiv 2024

  20. [20]

    K.et al.Atypical ductal hyperplasia: interobserver and intraobserver variability.Mod

    Jain, R. K.et al.Atypical ductal hyperplasia: interobserver and intraobserver variability.Mod. Pathol.24, 917–923, DOI: 10.1038/modpathol.2011.66 (2011)

    work page doi:10.1038/modpathol.2011.66 2011

  21. [21]

    H.et al.Understanding diagnostic variability in breast pathology: lessons learned from an expert consensus review panel.Histopathology65, 240–251, DOI: 10.1111/his.12387 (2014)

    Allison, K. H.et al.Understanding diagnostic variability in breast pathology: lessons learned from an expert consensus review panel.Histopathology65, 240–251, DOI: 10.1111/his.12387 (2014)

    work page doi:10.1111/his.12387 2014

  22. [22]

    arXiv preprint arXiv:2405.10254 (2024) 12 J

    Shaikovski, G.et al.Prism: A multi-modal generative foundation model for slide-level histopathology.arXiv preprint arXiv:2405.10254(2024). 12/16 3 Appendix We present representative samples from the datasets used in this study. Figures 4 and 5 are from two different patients in the 50-patient dataset, corresponding to the subtypes Invasive carcinoma, NST ...

    work page arXiv 2024