A Method for Normalizing Histology Slides for Quantitative Analysis
Macenko et al.
Metadata checked
Robustness, domain shift and privacy
Robustness is not one score. A pathology system can change because tissue was prepared differently, a scanner renders colour differently, disease prevalence changes, the case mix changes, annotation policy moves, or the relation between morphology and the target itself differs. Privacy is a separate contract: keeping slides at each institution is useful, but it does not by itself make shared model updates private.
1. Name the shift before choosing a remedy
Section titled “1. Name the shift before choosing a remedy”Use a specimen-aware causal sketch rather than the generic phrase “domain gap.” Record patient and disease factors; collection and fixation; stain batch; scanner, magnification, and compression; site and time; annotation policy; sampling; and the deployed endpoint. Then distinguish:
| Shift | What changes | What must be checked |
|---|---|---|
| Acquisition / covariate | Colour, focus, resolution, artifact, or morphology distribution | Scanner- and site-stratified performance; feature/site predictability; QC denominators |
| Prevalence / label | Class or outcome frequency | Calibration, predictive values, threshold utility, and subgroup denominators |
| Conditional / concept | The relation between morphology and target | Fresh labels, error adjudication, temporal and external validation; input-only monitoring cannot prove safety |
| Workflow | Case selection, users, decision process, or intervention | Intended-use protocol, human–AI interaction, turnaround time, overrides, and failure escalation |
Stain normalization, augmentation, domain adaptation, domain generalization, calibration, monitoring, and federated learning answer different questions. Never treat them as interchangeable checkboxes.
2. Measurement and intervention ladder
Section titled “2. Measurement and intervention ladder”Control the image pipeline
Section titled “Control the image pipeline”Macenko (2009) and Vahadane (2016) estimate stain representations and transform colour while trying to preserve morphology. Tellez et al. (2019) instead tests normalization and augmentation as learning interventions. These methods can reduce some colour variation, but no colour transform establishes transportability on its own; a transform may remove diagnostic chromatic evidence or leave non-colour site signals untouched.
Learn invariance without target labels
Section titled “Learn invariance without target labels”DANN (2016) gives the general domain-adaptation anchor: a feature extractor learns the task while confusing a source/target discriminator. That training contract normally sees unlabelled target inputs, so it is not a strict unseen-domain generalization experiment.
STRAP (2021) uses adaptive instance normalization with medically irrelevant artistic styles to disrupt texture and colour cues while retaining gross content. It reports stronger external CRC and CAMELYON17-WILDS results than the paper’s stain augmentation and stain normalization comparators. The evidence supports this tested recipe; it does not prove that every morphology-preserving style transform is safe.
Translate a preparation domain only with identity controls
Section titled “Translate a preparation domain only with identity controls”F2FLDM (2025) targets a different shift: frozen-section artifacts versus routine FFPE appearance. It learns frozen/FFPE pathology-feature distributions without spatially paired sections, interpolates a translated target embedding with the source embedding, uses DDIM inversion and LoRA-adapted latent diffusion, and sparsifies conditional noise changes with an L0 regularizer. On the paper’s held-out kidney cases, the FFPE-trained subtype classifier moves from frozen-section AUC/accuracy 81.99/61.97% to 94.26/80.34% with the regularized model; the unregularized model has slightly higher AUC but lower accuracy.
This is a preparation-translation intervention, not proof of pixelwise correction. Frozen and FFPE sections are different physical cuts; CaseFD, a downstream classifier, and a patch reader study constrain plausibility but cannot establish that an edited gland or nucleus is the same tissue. Preserve the source image, compare raw and translated tracks, split by case, report finding-level hallucinations, and validate new organs/sites before treating translation as a robustness remedy.
Benchmark natural shift
Section titled “Benchmark natural shift”WILDS (2021) fixes realistic distribution-shift datasets, including hospital shift in Camelyon17. Its central lesson is procedural: freeze metadata-defined in-distribution, validation, and out-of-distribution splits, and do not tune on the final OOD test set.
MIDOG (2023) tests mitosis detectors across scanner domains using containerized submissions, physical-distance matching, and scanner-stratified results. The top eligible final-test F1 was 0.748, while a post-challenge top-five ensemble reached 0.773 and must not be presented as an eligible single submission. The benchmark uses selected breast-cancer regions, not an end-to-end WSI diagnostic workflow.
Monitor a locked model
Section titled “Monitor a locked model”Fréchet Domain Distance (2023) applies a Fréchet distance to slide-level summaries of the top-attended positive CLAM features. In the reported breast lymph-node study, the K=64 positive-evidence variant correlated more strongly with MCC change than the compared confidence or entropy summaries. It is a shift signal, not a performance guarantee: a detected shift can be harmless or beneficial, and an undetected shift can still alter the label relation.
PathoROB (2026) and ZIO (2026) extend the operational view toward controlled perturbation testing, shift indices, uncertainty, and abstention. Monitoring needs an action policy: threshold, confirmatory labels, rollback or review path, owner, and a logged denominator.
VIPER supplies a different stress test: expert veterinary and toxicologic pathology questions over rat H&E regions probe transfer outside the human-pathology distributions that dominate VLM development. Its public evaluation code and dataset support repeatable audits, but 419 1,024-pixel ROIs are not WSIs and one species/toxicology setting is not a universal domain-shift benchmark. MVC-Bench adds calibration under medical modality and prompting changes; keep discrimination, calibration, domain shift, and clinical utility as separate outcomes.
3. Federated learning is not automatically private
Section titled “3. Federated learning is not automatically private”FedAvg (2017) establishes repeated local optimization and server aggregation. HistoFL (2022) adapts this pattern to gigapixel WSI by keeping tiles local, learning attention-MIL models, perturbing model weights with Gaussian noise, and averaging client updates. It reports that moderate tested noise scales retain much of the centralized retrospective result; however, the experiment does not report an end-to-end auditable privacy budget, secure aggregation, or protection against every update attack.
DP-SGD (2016) is the prerequisite for a formal differential-privacy claim: define adjacency, clip per-example contributions, add calibrated noise, and account for cumulative privacy loss. A Gaussian perturbation coefficient is not interchangeable with an explicit epsilon/delta guarantee.
Practical Secure Aggregation (2017) adds a different layer. Pairwise masks cancel inside the sum; threshold shares repair masks when clients drop; and a private self-mask plus a signed consistency round prevents a malicious server from manufacturing inconsistent dropout views. Under its stated corruption threshold, the server obtains an aggregate rather than each qualifying client’s plaintext update. This is cryptographic confidentiality, not differential privacy, poisoning resistance, availability, or a bound on what the aggregate and final model reveal.
Deep Leakage from Gradients (2019) shows why the update interface needs an explicit attack test. Given known weights and a directly observed small-batch gradient, it optimizes dummy inputs and labels until their parameter gradients match, using derivatives through the gradient calculation. The paper demonstrates image and text recovery in a controlled white-box setting, with modified twice-differentiable networks, batches up to eight, and images up to 64×64. Reproduce the attack against the actual object a pathology client releases—a per-step gradient, multi-step delta, compressed update, or secure aggregate. A failed DLG run is not a privacy guarantee, and the original attack’s success is not a universal risk estimate for every WSI federation.
HistoFS (2025) addresses non-IID federation by clustering patch-feature style statistics, sharing selected pseudo-style centroids, applying feature-space adaptive instance normalization with remote styles, and aligning attention as an ROI-preservation proxy. It improves reported RCC, HER2, and unseen Camelyon17 aggregates over HistoFL in the paper’s protocol and lowers the compared communication volume. Its conclusion explicitly leaves joint privacy protection for future work; data locality and style sharing must therefore not be labelled a privacy proof.
For any federated pathology study, report:
- The client unit, sampling, non-IID construction, participation rate, local epochs, optimizer state, and aggregation rule.
- Every object leaving a client: weights, gradients, optimizer state, feature moments, style centroids, metrics, logs, and identifiers.
- Threat model, attacker knowledge, adjacency definition, clipping, noise, accountant, epsilon/delta, secure aggregation, authentication, and revocation.
- Per-site and worst-site utility, calibration, communication, compute, failures, and comparison with local-only and centralized protocols.
4. Minimum external-validation protocol
Section titled “4. Minimum external-validation protocol”Before claiming robustness, freeze the model, preprocessing, checkpoint, threshold, and analysis plan. Split at patient level; describe specimen and slide hierarchy; retain site, scanner, stain batch, calendar time, and subgroup metadata; report confidence intervals and denominators; and include both aggregate and worst-group results. Test native, normalized, augmented, and calibrated tracks without mixing their claims. Audit overlap and repeated specimens. Adjudicate errors to determine whether failures arise from QC, sampling, representation, thresholding, label policy, or genuinely changed biology.
A practical deployment loop is:
- Baseline: lock the intended-use population, model, operating point, and external evidence.
- Observe: measure input/QC drift, output drift, uncertainty, abstention, and workflow change with complete denominators.
- Confirm: acquire representative labels when a signal fires; an unlabeled distance alone cannot establish performance loss.
- Act: review, recalibrate, retrain, roll back, or narrow intended use under a versioned change-control plan.
- Revalidate: repeat site/time/subgroup evaluation and privacy/security review after any material change.
Sparse features as a domain-shift audit
Section titled “Sparse features as a domain-shift audit”An SAE can help inspect whether an embedding encodes scanner or preparation information, but removing a detectable domain signal does not automatically preserve useful tissue biology. Start with the SAE learning guide, then distinguish the acquisition unit and external evaluation of each study.
Pathology concepts: distinguish representation from classifier reliance
Early pathology SAE work relates features to biological annotations and encoder depth. PATH-DISSECT adds depth-resolved selectivity checks and controlled interventions inside the frozen encoder, measured through a frozen linear probe. A represented tissue concept need not be used by every downstream head.
| Primary paper · year · evidence | Core idea and why to read it | Official resources |
|---|---|---|
| PATH-DISSECT: Interpreting Concept Representation and Classifier Reliance in Pathology Foundation Models with Sparse AutoencodersMayur Parvatikar, Xingjian Li, and Min Xu2026 · MICCAI MI4MedFM Workshop · acceptedOfficial MICCAI 2026 Satellite Events page supplies the 11-page camera-ready; Springer DOI and final pages were pending at inspection. This is MI4MedFM workshop work on tissue patches, not main-conference or patient-level WSI outcome validation. Causal claims concern interventions on a frozen encoder plus linear readout, not biological causation. | PATH-DISSECT separates a concept being encoded from a classifier relying on it. It trains sparse autoencoders at multiple blocks of frozen pathology transformers, validates class-associated features, then edits those features inside the forward pass while measuring a frozen linear probe. Its intervention is on model activations, not a causal experiment on tissue biology. | Official camera-ready PDF · Author implementation · Workshop OpenReview record |
Reading question: Is a tissue concept encoded, and does the specified classifier actually depend on it?
Evidence boundary: Tile-level annotations and linear-probe interventions are not WSI outcome validation or causal tissue biology. Preserve patch sampling, patient/site splits, feature-selection scope, and error-preserving controls.
Identify domain features without erasing task-relevant morphology
EXPOSE trains an SAE on pathology embeddings, identifies domain-associated latent coordinates, and masks selected coordinates before downstream prediction. It studies whether this intervention reduces acquisition-domain dependence while retaining useful outcome information.
| Primary paper · year · evidence | Core idea and why to read it | Official resources |
|---|---|---|
| EXPOSE: Explainable and Domain-Robust Embeddings from Pathology Vision Foundation Models using Sparse AutoencodersAnja Witte, Maximilian Lennartz, Jan Baumbach, Guido Sauter, Stefan Bonn, Patrick Fuhlert, and Marina Zimmermann2026 · ECCV eXCV Workshop · acceptedIndependently verified on the official eXCV 2026 Proceedings Track list, not ECCV main conference. Scientific review uses arXiv:2608.28191v1; a final typeset proceedings version was not verified. Evaluation is prostate tissue-microarray spots, not a whole-slide benchmark, and includes a thin-section failure case. | EXPOSE learns a sparse basis of frozen pathology features, identifies dimensions that distinguish two acquisition domains, and masks selected dimensions before training a relapse classifier. It is an automatic domain-feature selection strategy, not a new image encoder. The evidence concerns prostate tissue-microarray spots and controlled preparation/scanner variants, not unrestricted WSI deployment. | Official workshop accepted-paper list · Author implementation · Versioned full manuscript and appendix |
Reading question: Does domain-feature suppression preserve prediction in every held-out acquisition condition?
Evidence boundary: The inspected study uses prostate tissue-microarray spots, not a general full-WSI benchmark. Results vary by domain; removing detectable acquisition information does not prove that only nuisance was removed.
Complete maintained paper route
Section titled “Complete maintained paper route”28 canonical readings · 2009–2026· route reviewed through October 2, 2026.
This is the complete structured list for this chapter, ordered chronologically from earlier foundations toward current work. The explanatory ladder above highlights the shortest conceptual route; this list also includes important branches, benchmarks, and current endpoints. It is generated from the same deduplicated catalog used by the AI background registry and My Library, so a future catalog update appears here automatically.
Show all 28 readings in maintained order
Structure-Preserving Color Normalization and Sparse Stain Separation for Histological Images
Vahadane et al.
Metadata checked
Domain-Adversarial Training of Neural Networks
Yaroslav Ganin et al.
Metadata checked
Deep Learning with Differential Privacy
Martín Abadi et al.
Official artifacts: Google differential privacy
Metadata checked
Communication-Efficient Learning of Deep Networks from Decentralized Data
H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, and Blaise Agüera y Arcas
Metadata checked
Practical Secure Aggregation for Privacy-Preserving Machine Learning
Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone, H. Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, and Karn Seth
Official artifacts: Version-of-record manuscript · DOI
Metadata checked
Quantifying the Effects of Data Augmentation and Stain Color Normalization in Convolutional Neural Networks for Computational Pathology
Tellez et al.
Metadata checked
Deep Leakage from Gradients
Ligeng Zhu, Zhijian Liu, and Song Han
Official artifacts: Final proceedings PDF · Authors' implementation
Metadata checked
The Impact of Site-Specific Digital Histology Signatures on Deep Learning Model Accuracy and Bias
Howard et al.
Metadata checked
Learning Domain-Agnostic Visual Representation for Computational Pathology Using Medically-Irrelevant Style Transfer Augmentation
Rikiya Yamashita, Jin Long, Snikitha Banda, Jeanne Shen, and Daniel L. Rubin
Peer-reviewed journal article in IEEE Transactions on Medical Imaging 40(12):3945–3954; the linked arXiv manuscript is an author version.
Official artifacts: Author manuscript · Official code
Metadata checked
WILDS: A Benchmark of in-the-Wild Distribution Shifts
Pang Wei Koh et al.
Metadata checked
Federated learning for computational pathology on gigapixel whole slide images
Ming Y. Lu, Richard J. Chen, Dehan Kong, Jana Lipkova, Rajendra Singh, Drew F. K. Williamson, Tiffany Y. Chen, and Faisal Mahmood
Peer-reviewed Medical Image Analysis article 76:102298; DOI 10.1016/j.media.2021.102298.
Official artifacts: DOI · Author manuscript · Official code
Metadata checked
Mitosis domain generalization in histopathology images — The MIDOG challenge
Marc Aubreville et al.
Peer-reviewed Medical Image Analysis article 84:102699; published online 23 November 2022 and assigned to the 2023 volume.
Official artifacts: Author manuscript · Official challenge and data · Reference container
Metadata checked
Detecting Domain Shift in Multiple Instance Learning for Digital Pathology Using Fréchet Domain Distance
Milda Pocevičiūtė, Gabriel Eilertsen, Stina Garvin, and Claes Lundström
Peer-reviewed MICCAI 2023 proceedings paper, pages 157–167; its arXiv upload appeared in 2024 and does not change the venue year.
Official artifacts: MICCAI paper record · Author manuscript
Metadata checked
F2FLDM: Latent Diffusion Models with Histopathology Pre-Trained Embeddings for Unpaired Frozen Section to FFPE Translation
Man M. Ho, Shikha Dubey, Yosep Chong, Beatrice Knudsen, and Tolga Tasdizen
Official artifacts: Author preprint · Official project page
Metadata checked
Distilling Foundation Models for Robust and Efficient Models in Digital Pathology
Alexandre Filiot, Nicolas Dop, Oussama Tchita, Auriane Riou, Rémy Dubois, Thomas Peeters, Daria Valter, Marin Scalbert, Charlie Saillard, Geneviève Robin, and Antoine Olivier
Peer-reviewed MICCAI 2025 proceedings paper.
Official artifacts: Official model · Official robustness benchmark
Metadata checked
Physical Color Calibration of Digital Pathology Scanners for Robust Artificial Intelligence-Assisted Cancer Diagnosis
Xiaoyi Ji et al.
Metadata checked
HistoFS: Non-IID Histopathologic Whole Slide Image Classification via Federated Style Transfer with RoI-Preserving
Farchan Hakim Raswa, Chun-Shien Lu, and Jia-Ching Wang
Peer-reviewed CVPR 2025 proceedings paper, pages 30251–30260.
Official artifacts: Official project · Official code
Metadata checked
VIPER: An Expert-Curated Benchmark for Vision-Language Models in Veterinary Pathology
Luca L. Weishaupt, Simone de Brot, Javier Asin, Llorenç Grau-Roma, Nic G. Reitsam, Andrew H. Song, Dongmin Bang, Stefan T. Kaluziak, Long Phi Le, Jakob Nikolas Kather, Faisal Mahmood, and Guillaume Jaume
Acceptance independently verified 2026-10-08 through official NeurIPS poster 139742: exact title and matching 12-author list, linked OpenReview 7986DZ9rPe. Scientific review and artifact checks remain tied to arXiv:2608.26382v1 and the 2026-08-30 audit, not a refreshed full-text/code/data inspection. VIPER evaluates rat toxicologic regions rather than whole-slide images; it is an ROI benchmark, not a WSI benchmark.
Official artifacts: Official evaluation code and documentation · Official benchmark dataset · Official NeurIPS 2026 acceptance record
Metadata checked
MVC-Bench: Benchmarking Calibration of Medical Vision-Language Models
Ashshak Sharifdeen, Shihab Aaqil Ahamed, Ufaq Khan, Muhammad Akhtar Munir Sujair Ibrahim, Mohamed Rafeek Mareer Ahamed, Yutong Xie, Imran Razzak, and Muhammad Haris Khan
The arXiv record reports acceptance to Findings of EMNLP 2026. No final ACL Anthology record was available when checked on 30 August 2026, so claims were verified against arXiv v1 and the official repository.
Official artifacts: Official implementation
Metadata checked
Benchmarking Vision-Language Models for Automated Pathology Diagnosis and Report Generation
Yumi Lee, Harim Oh, Hyoryung Kim, Minji Kim, Eunsu Kim, Hyeseong Lee, Junya Fukuoka, Andrey Bychkov, Jijgee Munkhdelger, Rajiv Kumar Kaushal, Ayushi Sahay, Rajni Yadav, Bharathi Prabakaran, Sulen Sarioglu, Serdar Balcı, Ilknur Turkmen, Yuri Tolkach, Christian Harder, Julian Westerdorf, Reinhard Buettner, Audun Ljone Henriksen, Sepp De Raedt, Byung Hyun Lee, Sungjin Lim, Joohoon Lee, Gwanghyun Kim, Se Young Chun, Suryakant Singh, Saarthak Kapse, Prateek Prasanna, Kyung A Kim, Yousun Kang, Sehwan Yoo, Sungman Hong, Shubham Innani, Michael Feldman, Spyridon Bakas, Ujjwal Baid, Prasad Dutande, Suhas Gajare, Bhakti Baheti, Serkan Sökmen, Ece Tuğba Cebeci, Ahmet Halıcı, Musa Balcı, Kardelen Peçenek, Srividhya Sainath, Kyongseok Jang, Messi H. J. Lee, Noorul Wahab, Bodong Du, Jiaming Zhang, Qixiang Zhang, Jang-Hwan Choi, and Sangjeong Ahn
arXiv v1 submitted 1 September 2026 reports the results of the MICCAI REG 2025 challenge. No final peer-reviewed proceedings or journal record was verified on 2 September 2026. The approximately 10,500 institutional WSI-report pairs are not a generally downloadable public dataset; the released repositories expose benchmark information and team artifacts.
Official artifacts: Official challenge repository · Official winning-system repository · Official challenge site
Metadata checked
Reliable Benchmarking of Artifact Detection in Computational Pathology: A Reproducibility and Uncertainty Analysis
Konstantinos Moutselos and Ilias Maglogiannis
arXiv v1 submitted 31 August 2026. The authors released reproducibility deposits on Zenodo; no peer-reviewed record was verified on 2 September 2026.
Official artifacts: Official reproducibility deposit · Referenced benchmark artifact deposit
Metadata checked
KDTViT Knowledge Distillation, Transfer Learning and Transformer Based Deep Learning Framework for Efficient Histopathology Image Classification
Aarv Mankodi, Praveen Kumar Shukla, and Hitesh Tekchandani
Peer-reviewed open-access journal article 16:20721. A MobileViT-XS is sequentially adapted from 40× through 400× using the previous checkpoint as teacher; BreakHis consists of biopsy images rather than WSIs, and the same-capacity teacher/student design targets retention rather than model compression.
Official artifacts: Open-access article
Metadata checked
One Model to Magnify Them All: Efficient Scale-Invariant Histopathology via Conditional Normalization and Continuous Magnification Training
Agnieszka Florkowska, Henning Müller, and Marek Wodzinski
arXiv v1 submitted 10 August 2026; no peer-reviewed venue was verified on 2026-09-02. A single CNN is conditioned on physical pixel size and trained over a continuous magnification range; each prediction still consumes one patch scale and the method is neither MIL nor knowledge distillation.
Official artifacts: Official code
Metadata checked
EXPOSE: Explainable and Domain-Robust Embeddings from Pathology Vision Foundation Models using Sparse Autoencoders
Anja Witte, Maximilian Lennartz, Jan Baumbach, Guido Sauter, Stefan Bonn, Patrick Fuhlert, and Marina Zimmermann
Independently verified on the official eXCV 2026 Proceedings Track list, not ECCV main conference. Scientific review uses arXiv:2608.28191v1; a final typeset proceedings version was not verified. Evaluation is prostate tissue-microarray spots, not a whole-slide benchmark, and includes a thin-section failure case.
Official artifacts: Official workshop accepted-paper list · Author implementation · Versioned full manuscript and appendix
Metadata checked
PATH-DISSECT: Interpreting Concept Representation and Classifier Reliance in Pathology Foundation Models with Sparse Autoencoders
Mayur Parvatikar, Xingjian Li, and Min Xu
Official MICCAI 2026 Satellite Events page supplies the 11-page camera-ready; Springer DOI and final pages were pending at inspection. This is MI4MedFM workshop work on tissue patches, not main-conference or patient-level WSI outcome validation. Causal claims concern interventions on a frozen encoder plus linear readout, not biological causation.
Official artifacts: Official camera-ready PDF · Author implementation · Workshop OpenReview record
Metadata checked
Towards Robust Foundation Models for Digital Pathology
Jonah Kömen et al.
Metadata checked
Training-Free Out-of-Distribution Detection for Pathology Whole-Slide Images
Sabri Mustafa Kahya et al.
The arXiv record describes the manuscript as under review; this catalog retains preprint status.
Official artifacts: Official code
Metadata checked