Enhancing Whole Slide Image Classification through Supervised Contrastive Domain Adaptation

Il\'an Carretero; Pablo Meseguer; Roc\'io del Amor; Valery Naranjo

REVIEW 4 major objections 5 minor 2 cited by

Enhancing Whole Slide Image Classification through Supervised Contrastive Domain Adaptation

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Adding a cross-domain constraint to supervised contrastive learning lifts cross-hospital whole-slide classification from 0.51–0.54 balanced accuracy to 0.88–0.93.

desk verdict Large empirical win for cross-hospital WSI classification, but the claimed cross-center constraint is not shown to be the cause — needs an ablation before the mechanism is credible. read the letter →

arxiv 2412.04260 v1 pith:5JXEBHDD submitted 2024-12-05 cs.CV cs.AI

Ilán Carretero , Pablo Meseguer , Rocío del Amor , Valery Naranjo This is my paper · ORCID

classification cs.CVcs.AI

keywords wholeslideimageclassificationdomainadaptationsupervisedcontrastivelearninghistopathologyskincancerfew-shotfoundationmodelembeddingsmultipleinstance

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

The paper proposes Supervised Contrastive Domain Adaptation (SCDA) to fix a practical obstacle in computational pathology: the same tumor type looks different when scanned and stained at different hospitals, so a model trained at one center often fails at another. SCDA takes slide-level embeddings from a frozen foundation model, PLIP, and trains a lightweight contrastive transformation that pulls same-class slides from the two hospitals together while pushing different classes apart. The central evidence is that balanced accuracy on unseen-hospital test sets rises from 0.54 and 0.51 without adaptation to 0.88 and 0.93 with SCDA, and that a few-shot version reaches comparable performance with roughly 8–10 labeled slides per class from the target hospital. If this holds, SCDA offers a low-cost alternative to stain normalization and heavy re-training for multi-center histopathology classification.

What carries the argument

The machinery is supervised contrastive learning with a cross-domain batch constraint, applied to bag-level embeddings produced by PLIP patch features followed by non-trainable global average pooling (BGAP). The loss rewards closeness between same-class slides from different centers and penalizes closeness to other classes, so the learned transformation projects both hospitals' representations into a common space; classification is then done by MI-SimpleShot, which compares each transformed slide to class prototypes and takes the largest similarity.

What would settle it

Run SCDA on the same dataset but with training batches that each contain only one center (so the cross-domain constraint is never satisfied), then measure balanced accuracy on the held-out center; if the 0.88/0.93 results still appear, the reported improvement is not caused by the proposed constraint.

Watch

Extended reading notes

Core claim

The paper's claim is that a single training constraint converts the raw embeddings of a frozen pathology foundation model into a domain-aligned slide representation. The constraint requires that every training batch contains, for each class, at least one sample from each of the two centers; the supervised contrastive loss then treats same-class cross-center samples as positives, making the model erase center-specific staining and scanning signals while preserving class identity. Applied to PLIP patch features aggregated by global average pooling, this produces transformed embeddings $C$ that are classified by nearest-class-prototype matching rather than a trained softmax head. In the paper's experiments, SCDA raises balanced accuracy on the held-out center from 0.51–0.54 to 0.88–0.93, and also outperforms Macenko stain normalization, which drops accuracy to 0.60–0.73.

Load-bearing premise

The cross-domain constraint only helps if every training batch actually contains labeled slides of every class from both centers, and the paper does not describe the batching procedure that guarantees this for rare classes or small few-shot sets.

Editorial extensions

If this is right

A hospital with an existing frozen patch-embedding model can adapt to a new center using only the slide-level embeddings and labels, without retraining the patch encoder.
Few-shot adaptation with around 8–10 labeled slides per class from the target center reaches accuracy comparable to training on the full target dataset.
Stain normalization, which can remove diagnostically useful color information, is not needed when SCDA is applied.
SCDA requires the same set of classes to be present in both hospitals, so it applies to tasks where the label space is shared across centers.

Reading between the lines

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

The paper does not test SCDA on more than two centers; if the same constraint is applied pairwise, the method may extend to multi-site cohorts, but the loss would then need to see several domains per batch rather than two.
The mechanism suggests a general recipe for removing scanner- and protocol-specific bias from any frozen embedding: enforce cross-batch same-class alignment during contrastive fine-tuning, not just data augmentation.
A direct test of the claimed mechanism would be to keep the loss but violate the cross-domain constraint in one class; if that class still improves, the gain comes from contrastive clustering rather than from domain alignment.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Here's the short version: the paper reports a big empirical win — balanced accuracy on cross-hospital skin WSI classification jumps from ~0.54/0.51 with frozen PLIP embeddings to 0.88/0.93 after a cheap supervised-contrastive adaptation step. If that result holds up, it's a useful practical trick. But the paper does not actually verify the mechanism it claims: the cross-center constraint is never shown to be the thing driving the gain.

What's new: the explicit cross-center batch constraint — forcing each training batch to contain both centers' examples of every class during supervised contrastive learning — is a legitimate extension of SupCon to domain adaptation in pathology. The authors do the right baseline comparisons (no adaptation, Macenko normalization), evaluate under a few-shot schedule with five seeds, and are upfront about the main limitations: labels required, matched class sets, only two hospitals. The method is low-compute and the t-SNE figures support the story that the adapted embeddings mix the centers.

The soft spots are real. First, Eq. (1) is standard SupCon; the constraint lives entirely in how batches are composed. Section 3.2 never states the batching scheme, the batch size, or whether the constraint was actually satisfiable for the rare classes (dermatofibrosarcoma has ~14 HCUV training slides — a random batch of 32 would often miss it). Second, and more important, there is no ablation comparing SCDA to vanilla SupCon on randomly mixed batches. Without that, the observed gain could just be standard SupCon clustering, with the constraint being inert. Third, the transformation T is underspecified — no architecture, optimizer, learning rate, epochs, or temperature — so the method isn't reproducible and we can't judge overfitting on ~486 training slides. Fourth, Table 2 reports no variance or significance despite five seeds, so we don't know how stable 0.88/0.93 is.

I'd slightly push back on the stress-test note's insistence that the constraint is unverifiable: if the authors enforced it, it's a real implementation detail, but the paper gives no evidence either way. The concern lands, just not as hard as "moot."

Who gets value: this is squarely for the pathology-ML community and anyone working on cheap adaptation of frozen foundation models. The empirical result is plausible and potentially important. But as written, the paper is a solid workshop-level finding that needs a revision before it can carry the load: add the ablation, specify the training pipeline, report the seed spread. I'd send it to peer review — it deserves referee time — but I'd hold off citing it myself until the attribution question is answered.

Referee Report

4 major / 5 minor

Summary. The paper proposes Supervised Contrastive Domain Adaptation (SCDA), a domain adaptation method for whole-slide image (WSI) classification. The method extracts patch-level features with the frozen PLIP foundation model, aggregates them via BGAP, and then learns a transformation of the slide-level embeddings using the supervised contrastive loss of Khosla et al. with an additional cross-domain constraint. The constraint is stated as: for every class S and every training batch B, the batch must contain at least one sample of class S from each of the two centers. Classification is then performed with MI-SimpleShot, a prototype-based classifier. The authors evaluate SCDA on a two-center skin cancer dataset (HCUV and HUSC) with six subtypes. Table 2 reports large balanced-accuracy gains for cross-hospital inference: for testing on HCUV after training on HUSC, BACC rises from 0.54 (PLIP alone) to 0.88 (PLIP+SCDA); for testing on HUSC after training on HCUV, from 0.51 to 0.93. When both centers are used for training, SCDA achieves 0.91 BACC on the merged test set, compared to 0.76 for PLIP alone and 0.68 for Macenko normalization. The paper also reports few-shot experiments (k = 2 to 10 samples per class) showing improved cross-center performance as the number of shots increases.

Significance. If the reported improvements hold, SCDA would be a simple and computationally efficient supervised domain adaptation method for histopathology: it requires no retraining of the feature extractor, no stain normalization, and it appears to work with only a few labeled examples from the target center. The use of a frozen foundation model and BGAP aggregation makes the method lightweight, and the few-shot evaluation is practically relevant for multi-center clinical deployment. The paper also provides qualitative t-SNE visualizations showing that SCDA reduces the center-induced separation in the embedding space. However, the scientific significance currently hinges on three unverified points: (i) the training procedure for the transformation T is unspecified, (ii) the cross-domain constraint is not shown to be satisfied in practice, and (iii) no ablation isolates the effect of the constraint from the effect of supervised contrastive learning itself. Because the proposed method's only novel ingredient is the constraint, these omissions prevent the reader from attributing the large reported gains to the claimed mechanism.

major comments (4)

[Section 3.2] The training procedure for the transformation T is not reported. The paper never specifies the batch size, the batching strategy, the number of epochs, the optimizer, the learning rate, the temperature τ, or how the cross-domain constraint is enforced in the optimization. Since Eq. (1) is the standard supervised contrastive loss, the constraint is the only novel component; without these details the reader cannot reproduce the method or verify that the reported gains are caused by the constraint.
[Section 2.3, Table 2] No ablation isolates the cross-domain constraint. The authors compare SCDA against PLIP without adaptation and against Macenko normalization, but not against vanilla supervised contrastive learning on random mixed-center batches, nor against supervised contrastive learning trained per center. Because Eq. (1) is the standard SupCon loss, the improvement from 0.54/0.51 to 0.88/0.93 could be due to supervised contrastive learning itself rather than the proposed constraint. This is a load-bearing point for the central claim, and it must be tested directly.
[Section 2.3, Figure 2] For rare classes and small shot counts, the constraint may be unsatisfiable. The class dfs has only 17 slides in HCUV, which after an 80/20 split leaves roughly 14 training slides. In the few-shot scenario with k = 2 HCUV shots per class, it is impossible for a batch to contain both HCUV and HUSC samples of every class unless the batch is constructed with heavy resampling or the constraint is relaxed. The paper does not state how such cases are handled, and if the constraint is not actually enforced during training, the described mechanism is not what produced the results.
[Table 2, Figure 2] No variance or significance information is provided for the main results. The few-shot experiments are run with five seeds, but Table 2 appears to report single numbers without error bars, and no statistical test is applied. Given that the improvements are the central evidence, the authors should report standard deviations or confidence intervals for the BACC values, especially for the comparison with the no-adaptation baseline.

minor comments (5)

[Eq. (1)] The displayed equation has typographical errors: the numerator repeats 'zizizi · zpzpzp', and the denominator 'zazaza' is not properly typeset. In addition, the definition of A(i) as 'all representations of the batch' is ambiguous, because in supervised contrastive learning the positive set for sample i should exclude i itself and include only same-class samples; this should be clarified.
[Table 2] The column and row labels are ambiguous. For example, the row 'HCUV' under 'Training centers' could mean training on HCUV alone, but the testing columns are labeled 'HCUV', 'HUSC', and 'HCUV∪HUSC'; the reader has to infer which rows correspond to same-center versus cross-center evaluation. A more explicit notation (e.g., 'train: HCUV, test: HUSC') would improve readability.
[Section 4.1, Figure 2] The '0-shot' starting point is not defined in the text. It appears to refer to using PLIP+BGAP features without any adaptation, but this should be stated explicitly, and the same applies to the horizontal dotted lines representing 'trained with both databases without using SCDA.'
[Section 5] There are several typos: 'contative' should be 'contrastive' in the conclusion, and 'HUCV' appears multiple times in Section 4.1 where 'HCUV' is intended. These should be corrected.
[Section 2.4] The definition of MI-SimpleShot is brief; in particular, the construction of the class prototypes W and the equation Ŷ = argmax(W^T C) assume that C and W are in the same space, but the paper does not state whether the prototypes are computed from the transformed embeddings C or from the original Z. This should be clarified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports an empirical benchmark with no derivation whose conclusions reduce to its inputs.

full rationale

The paper makes no analytic derivation whose conclusions are equivalent to its premises. Equation (1) is the standard supervised contrastive loss, and the cross-domain constraint is a stated batch-composition condition; the reported gains in Table 2 and Figure 2 are empirical measurements on a fixed held-out test split of the two centers. The transformation T is underspecified and the batching scheme is not described, and no ablation isolates the constraint from vanilla SupCon on mixed-center batches; these are reproducibility and attribution concerns, not circularity. The authors' own dataset citation [7] includes a coauthor, but it is used as a data source, not as a load-bearing justification of the method. Evaluation on the same two centers used for training limits external validity but does not make the result circular. No fitted parameter is renamed as a prediction, and no self-citation chain forces the reported conclusion. Therefore no circular step can be exhibited, and the appropriate finding is no significant circularity.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No new physical entities or fitted constants are introduced. The method has two unstated training hyperparameters and four modeling assumptions, the most fragile being the batch-level cross-domain constraint.

free parameters (2)

Supervised contrastive temperature tau = not reported
Scales similarity in Eq. (1); affects clustering tightness; no value is given in the experimental setup.
SCDA training hyperparameters (optimizer, learning rate, epochs, batch size) = not reported
The transformation T is trained but no training configuration is provided; batch size determines whether the cross-domain constraint can be satisfied.

assumptions (4)

domain assumption Both centers share the same six classes and labels are available for all training slides
Stated in the problem formulation and in the limitations; the method cannot handle new classes or unlabeled centers.
standard math WSI-level labels apply to all constituent patches (MIL assumption)
The bag-level representation averages patch features; this is the standard multiple instance learning assumption in pathology.
domain assumption PLIP features contain class-discriminative but center-biased information that can be aligned by a transformation
The method relies on frozen PLIP features and BGAP; if the domain shift destroyed class information, contrastive alignment could not recover it.
ad hoc to paper Every training batch can satisfy the cross-domain constraint for every class
The constraint in Section 2.3 requires both centers represented for each class in each batch; this is assumed without specifying batching or addressing rare classes.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Enhancing Whole Slide Image Classification through Supervised Contrastive Domain Adaptation." pith.science (2026). https://pith.science/paper/5JXEBHDD

@misc{pith2026241204260,
  author       = {Pith},
  title        = {Pith review of: Enhancing Whole Slide Image Classification through Supervised Contrastive Domain Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5JXEBHDD}},
  note         = {Machine review of arXiv:2412.04260}
}

read the original abstract

Domain shift in the field of histopathological imaging is a common phenomenon due to the intra- and inter-hospital variability of staining and digitization protocols. The implementation of robust models, capable of creating generalized domains, represents a need to be solved. In this work, a new domain adaptation method to deal with the variability between histopathological images from multiple centers is presented. In particular, our method adds a training constraint to the supervised contrastive learning approach to achieve domain adaptation and improve inter-class separability. Experiments performed on domain adaptation and classification of whole-slide images of six skin cancer subtypes from two centers demonstrate the method's usefulness. The results reflect superior performance compared to not using domain adaptation after feature extraction or staining normalization.

Figures

Figures reproduced from arXiv: 2412.04260 by the authors.

**Figure 1.** Method overview. In this article, we address inter-center domain shifts through supervised contrastive learning. Concretely, a constraint is introduced in the model training by forcing samples of the same class to be closely spaced. The inclusion of this condition strengthens inter-class clustering and removes inter-center variability. arXiv:2412.04260v1 [cs.CV] 5 Dec 2024 [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** In-depth study of SCDA under the few-shot learning paradigm. In (a) HUSC train set is used, and HCUV samples are added. In (b), the reverse procedure is evaluated. The results of the HUSC test set are shown in blue and those of HCUV in orange. The starting point on the curves refers to the 0-shot case, where only the PLIP and BGAP feature extractors are used to predict. The final triangle in the curves refers to th… view at source ↗

**Figure 3.** 2D t-SNE representations of feature embeddings extracted by PLIP and aggregation via BGAP. The subfigures illustrate (a) Representation for the embeddings without any added processing. (b) Representation of the embeddings obtained with PLIP and BGAP by previously applying the Macenko normalization method. (c) Representation of the embeddings by applying SCDA. 5. Conclusion The intrinsic staining and digitization v… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

HASD: Hierarchical Adaption for pathology Slide-level Domain-shift
cs.AI 2025-06 conditional novelty 6.0 of 10

A hierarchical domain adaptation framework improves slide-level HER2 grading and survival prediction across medical centers by aligning features at domain, slide, and patch levels.
MorphGen: Morphology-Guided Representation Learning for Robust Single-Domain Generalization in Histopathological Cancer Classification
cs.CV 2025-08 conditional novelty 5.0 of 10

MorphGen uses supervised contrastive learning to align histopathology images with nuclear masks and applies SWA, reporting improved out-of-domain cancer classification accuracy on CAMELYON17, BCSS, and OCELOT.

Reference graph

Works this paper leans on

8 extracted references · 7 canonical work pages · cited by 2 Pith papers

[1]

Mea- suring domain shift for deep learning in histopathology,

K. Stacke, G. Eilertsen, J. Unger, and C. Lundstr ¨om, “Mea- suring domain shift for deep learning in histopathology,” IEEE journal of biomedical and health informatics , vol. 25, no. 2, pp. 325–336, 2020

work page 2020
[2]

A method for normalizing histology slides for quantitative analysis,

M. Macenko, M. Niethammer, J. S. Marron, D. Borland, J. T. Woosley, X. Guan, C. Schmitt, and N. E. Thomas, “A method for normalizing histology slides for quantitative analysis,” in 2009 IEEE international symposium on biomedical imaging: from nano to macro, pp. 1107–1110, IEEE, 2009

work page 2009
[3]

Unpaired image-to-image translation using cycle-consistent adversarial networks,

J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” inProceedings of the IEEE international conference on computer vision, pp. 2223–2232, 2017

work page 2017
[4]

Distill- soda: Distilling self-supervised vision transformer for source- free open-set domain adaptation in computational pathology,

G. Vray, D. Tomar, B. Bozorgtabar, and J.-P. Thiran, “Distill- soda: Distilling self-supervised vision transformer for source- free open-set domain adaptation in computational pathology,” IEEE Transactions on Medical Imaging, 2024

work page 2024
[5]

Supervised contrastive learning,

P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y . Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learning,” Advances in neural information processing systems, vol. 33, pp. 18661–18673, 2020

work page 2020
[6]

A general-purpose self-supervised model for computational pathology,

R. J. Chen, T. Ding, M. Y . Lu, D. F. Williamson, G. Jaume, B. Chen, A. Zhang, D. Shao, A. H. Song, M. Shaban, et al., “A general-purpose self-supervised model for computational pathology,”arXiv preprint arXiv:2308.15474, 2023

arXiv 2023
[7]

Annotation protocol and crowdsourcing multiple instance learning classification of skin histological images: The cr-ai4skin dataset,

R. Del Amor, J. P ´erez-Cano, M. L ´opez-P´erez, L. Terradez, J. Aneiros-Fernandez, S. Morales, J. Mateos, R. Molina, and V . Naranjo, “Annotation protocol and crowdsourcing multiple instance learning classification of skin histological images: The cr-ai4skin dataset,”Artificial Intelligence in Medicine, vol. 145, p. 102686, 2023

work page 2023
[8]

A visual–language foundation model for pathology im- age analysis using medical twitter,

Z. Huang, F. Bianchi, M. Yuksekgonul, T. J. Montine, and J. Zou, “A visual–language foundation model for pathology im- age analysis using medical twitter,” Nature medicine, vol. 29, no. 9, pp. 2307–2316, 2023

work page 2023

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [1]

Mea- suring domain shift for deep learning in histopathology,

K. Stacke, G. Eilertsen, J. Unger, and C. Lundstr ¨om, “Mea- suring domain shift for deep learning in histopathology,” IEEE journal of biomedical and health informatics , vol. 25, no. 2, pp. 325–336, 2020

work page 2020

[2] [2]

A method for normalizing histology slides for quantitative analysis,

M. Macenko, M. Niethammer, J. S. Marron, D. Borland, J. T. Woosley, X. Guan, C. Schmitt, and N. E. Thomas, “A method for normalizing histology slides for quantitative analysis,” in 2009 IEEE international symposium on biomedical imaging: from nano to macro, pp. 1107–1110, IEEE, 2009

work page 2009

[3] [3]

Unpaired image-to-image translation using cycle-consistent adversarial networks,

J.-Y . Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” inProceedings of the IEEE international conference on computer vision, pp. 2223–2232, 2017

work page 2017

[4] [4]

Distill- soda: Distilling self-supervised vision transformer for source- free open-set domain adaptation in computational pathology,

G. Vray, D. Tomar, B. Bozorgtabar, and J.-P. Thiran, “Distill- soda: Distilling self-supervised vision transformer for source- free open-set domain adaptation in computational pathology,” IEEE Transactions on Medical Imaging, 2024

work page 2024

[5] [5]

Supervised contrastive learning,

P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y . Tian, P. Isola, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learning,” Advances in neural information processing systems, vol. 33, pp. 18661–18673, 2020

work page 2020

[6] [6]

A general-purpose self-supervised model for computational pathology,

R. J. Chen, T. Ding, M. Y . Lu, D. F. Williamson, G. Jaume, B. Chen, A. Zhang, D. Shao, A. H. Song, M. Shaban, et al., “A general-purpose self-supervised model for computational pathology,”arXiv preprint arXiv:2308.15474, 2023

arXiv 2023

[7] [7]

Annotation protocol and crowdsourcing multiple instance learning classification of skin histological images: The cr-ai4skin dataset,

R. Del Amor, J. P ´erez-Cano, M. L ´opez-P´erez, L. Terradez, J. Aneiros-Fernandez, S. Morales, J. Mateos, R. Molina, and V . Naranjo, “Annotation protocol and crowdsourcing multiple instance learning classification of skin histological images: The cr-ai4skin dataset,”Artificial Intelligence in Medicine, vol. 145, p. 102686, 2023

work page 2023

[8] [8]

A visual–language foundation model for pathology im- age analysis using medical twitter,

Z. Huang, F. Bianchi, M. Yuksekgonul, T. J. Montine, and J. Zou, “A visual–language foundation model for pathology im- age analysis using medical twitter,” Nature medicine, vol. 29, no. 9, pp. 2307–2316, 2023

work page 2023