acceptodds
Under review as a conference paper at ICLR 2027

SlideTTT: Slide-Specific Test-Time Training for Generalizable Whole-Slide Images Analysis

Abstract

Multiple instance learning (MIL) has become the dominant paradigm for weakly supervised whole-slide image (WSI) analysis, where large-scale WSIs are divided into patches and their representations are subsequently aggregated for slide-level prediction. However, most existing MIL methods still follow a static inference paradigm, applying a fixed aggregator learned during training to all test WSIs. In the presence of substantial inter-slide heterogeneity and cross-center distribution shifts, such a unified aggregation strategy may limit model generalization. To address this issue, we propose SlideTTT, a slide-specific test-time training framework that enables each WSI to adapt its aggregation process before prediction. At its core, we introduce a Slide TTT Layer that constructs an inner-learning problem from patch tokens, where key-value pairs dynamically update lightweight fast weights and query tokens retrieve slide-specific representations from the adapted state. The layer combines complementary global and local adaptation: the global branch captures long-range semantic dependencies, while the local branch models short-range spatial interactions through convolutional fast weights. Moreover, the Slide TTT Layer is designed as a modular component that can be readily integrated into other Transformer-based MIL architectures. We evaluate SlideTTT on five pathology tasks using paired internal and external cohorts. SlideTTT achieves the highest average accuracy across internal, external, and overall evaluations, with more pronounced gains on external cohorts, demonstrating the effectiveness of slide-specific adaptation for improving generalization to unseen data. Code is available at https://anonymous.4open.science/r/SlideTTT.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Multiple Instance Learning, Test-Time Training, Slide-Specific Adaptation, Dynamic Aggregation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_pzmcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SlideTTT, a slide-specific test-time training framework for whole-slide image analysis. The Slide TTT Layer, which forms the core of SlideTTT, is configured to combine complementary global and local adaptation, where the global branch captures long-range semantic dependencies and the local branch models short-range spatial interactions through convolutional fast weights. In the experiments in the paper, it is reported that the proposed method shows superior results compared with conventional methods.

Critical strengths

As far as its model structure is concerned, SlideTTT presented in this paper is expected to be applicable to data other than WSIs. This point is considered to be a strength from the perspective of the generality of the proposed method.

In addition, a very large number of experimental results are presented, and I was able to sufficiently understand the effectiveness of the proposed method.

Critical weaknesses

Although this study proposes SlideTTT targeting WSIs, doubts remain as to whether it can be said to be a study specialized for tasks targeting WSIs. This is because SlideTTT itself can also be considered a general purpose method and is considered to be applicable to various fields. As mentioned above, this is undoubtedly a strength of this study, but on the other hand, it can also be said that the reason for limiting the target to WSIs in this paper is unclear.

The explanation in Section 3 is confined to content regarding the structure and processing of the model, and it cannot be said that why the proposed method comes to solve the problems that have existed conventionally is clearly described. It is desirable that it be clearly shown whether these model structures were derived heuristically, or whether they were introduced with confidence based on prior knowledge.

Most of the experimental results report the accuracy of the model, but it would have been desirable for it to be made clear what tendencies the accuracy of the model shows with respect to the statistical properties of the target data.

Questions for authors

As described in both the Strengths and the Weaknesses, this study can also be said to be a general purpose study. The reviewer would like the authors to clarify where the reason lies that it must be WSIs. In addition, if it is a general purpose effort that does not need to be specialized for WSIs, The reviewer thinks that the explanation of this paper should originally be written in that way, and that the evaluation also does not need to be limited to WSIs.

In this paper, it is stated that, in order to perform a more rigorous evaluation of cross-center generalization, the smaller cohort was consistently used for model development and the larger cohort for external evaluation. The reviewer would like the authors to show whether the same conclusion can be obtained even when this is reversed.

It is clearly shown that the accuracy of the proposed method is superior to that of conventional methods. On the other hand, the reviewer hopes that it will also be clearly shown, with respect to the tendencies of the data, in what cases superior results are obtained and in what cases difficulties are involved.

Reviewer_3K6KOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes SlideTTT, a test-time training framework for WSI-level multiple instance learning that enables slide-specific adaptation during aggregation. The core part is SlideTTT layer, which combines complementary global and local adaptation: the global branch captures long-range semantic dependencies, while the local branch models short-range spatial interactions through convolutional fast weights.

Critical strengths

How to solve substantial inter-slide heterogeneity and cross-center distribution shifts is an important problem in digital pathology area. The methods part is clear to understand with code.

Critical weaknesses
  1. Motivation and method do not match. The paper says it wants to solve substantial inter-slide heterogeneity and cross-center distribution shifts. However, the method treats TTT as a sequence modeling mechanism and this mechanism can't solve substantial inter-slide heterogeneity and cross-center distribution shifts. The paper provides no evidence linking the mechanism to cohort shift.
  2. Novelty is limited. The global and local branches of SlideTTT layer are the same as in the ViTTT paper [1]. This paper just replaces the input with WSI patch embeddings. The whole framework is transMIL [2], and this paper only replaces transMIL’s attention with the ViTTT layer.
  3. The experiment is insufficient. Following the introduction, if this paper wants to solve the domain shift issue, it should compare with domain adaptation methods, such as DANN, stain normalization/augmentation, and other test time adaptation methods.
  4. Paper mentions "Existing MIL methods rely on fixed aggregators for all test WSIs". However, attention-based MIL will change attention weight by the input samples, which is same with SlideTTT. And SlideTTT's weights are also fixed, and it learns function of the input: the same single inner step is executed for every training slide and every test slide.

[1] Han, Dongchen, et al. "ViT 3^ 3: Unlocking Test-Time Training in Vision." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026. [2] Shao, Zhuchen, et al. "Transmil: Transformer based correlated multiple instance learning for whole slide image classification." Advances in neural information processing systems 34 (2021): 2136-2147.

Questions for authors

What's the motivation of this paper, solving domain shift in WSI or using TTT as a sequence modeling mechanism for WSI analysis? What is the difference between slideTTT layer and ViTTT layer? SlideTTT is a linear-attention-type layer, how can it perform “test-time updates”?

Reviewer_9zQmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SlideTTT presents slide-specific test-time adaptation for whole-slide image classification. An unlabeled key–value objective updates lightweight fast weights in global gated and local convolutional branches, enabling aggregation to rely on the given slide. The resulting architecture employs two TTT layers divided by multiscale positional encoding. Experiments use frozen UNI features over five classification tasks, along with component ablations and integration into DGR and MHIM.

Critical strengths
  1. Slide-specific adaptation addresses an important generalization problem.

    Variations in tissue composition and acquisition conditions may make aggregation learned on one cohort less effective on another. SlideTTT offers a useful way to condition aggregation on the given slide without requiring its label or keeping adaptation state across patients. External ER and PR gains in both accuracy and AUC provide promising evidence that this approach can help cross-cohort prediction.

  2. The experiments demonstrate useful application breadth and initial evidence of modularity.

    Evaluation covers breast, colorectal, and lung tasks, while shared frozen UNI features help control the representation employed by competing aggregators. Experiments with DGR and MHIM also show that the proposed layer can be used beyond the authors’ standalone architecture. Although the gains differ across tasks and metrics, these results make the approach relevant to several existing MIL pipelines.

Critical weaknesses
  1. The comparisons do not adequately isolate the value of the proposed adaptation.

    ViT³ already offers strongly related gated/convolutional inner models and an update procedure, whereas MDPE closely follows TransMIL’s positional encoder. A matched ViT³-based MIL model is thus required to determine what the SlideTTT modifications contribute. The no-update ablation does not settle this issue: under the described architecture, removing global updates seems to remove the route through which patch information reaches CLS. Its weak performance may consequently reflect loss of aggregation rather than the specific benefit of adaptation.

    The morphology motivation additionally requires better positioning against existing input-dependent representations: fixed learned parameters do not imply input-independent aggregation. R2T-MIL, DTFD-MIL [2], and KAT [3] are missing from the experiments despite offering relevant alternatives through feature re-embedding, feature distillation, and spatial aggregation. Matched comparisons would help determine whether TTT provides benefits beyond these mechanisms.

    Foundation models and their related slide representations need a more direct comparison. Although the paper contains UNI features with ABMIL—the conventional weakly supervised setting evaluated within the UNI paper—it does not contrast with HIPT’s pretrained hierarchical representation [5] or PANTHER’s slide-specific morphology representation [1]. Such alternatives are especially relevant to the claim that test-time adaptation is required to accommodate tissue heterogeneity. UNI’s MI-SimpleShot is another absent comparison, though its use of annotated training ROIs makes it an additional experiment requiring explicitly matched supervision rather than a mandatory slide-label-only baseline.

    Standalone mean, max, log-sum-exp (LSE), and MaxSoft [4] pooling are important controls for the practical value of the extra adaptation machinery. Table 4’s CLS-plus-mean variant keeps TTT and thus does not address this question. Collectively, these omissions leave the contribution insufficiently distinguished from both closely related architectures and simpler aggregation alternatives.

  2. The reliability and scope of the generalization benefit remain insufficiently established.

    AUC is present in the appendix, yet the primary discussion focuses accuracy without sufficiently reconciling the different metrics. External accuracy gains over the best reported baseline are just 0.22–0.35 percentage points for PAM50, MSI, and NSCLC. For MSI, 76.38% accuracy accompanies 50.75% balanced accuracy and 59.18 ± 20.0% AUC, constraining the evidence for useful discrimination. Plug-in effects are likewise mixed: DGR’s external average accuracy declines while its AUC and balanced accuracy improve. These results warrant task- and metric-specific conclusions rather than a generally positive account of transfer.

    Five-seeds do not establish uncertainty in paired performance differences on the external patient population. Patient-disjoint internal splitting is not reported, and inconsistent external PR F1 and balanced-accuracy values between the main-comparison and ablation tables further complicate interpretation. These issues matter especially when the reported improvements are small.

    All studies employ one frozen encoder, UNI, leaving uncertain whether the improvements rely on pathology-specific pretraining. Evaluating another pathology encoder and an ImageNet-pretrained ResNet-50 could help differentiate a broadly useful aggregation mechanism from a benefit specific to the chosen representation. The assessment also lacks a metastasis-detection transfer setting such as CAM16-to-CAM17, which would test generalization when diagnostic tissue occupies only a small fraction of the slide. These limitations limit the scope of the claims, while the existing cross-cohort results remain relevant evidence.

References [1] Song et al., “Morphological Prototyping for Unsupervised Slide Representation Learning in Computational Pathology”.

[2] Zhang et al., “DTFD-MIL: Double-Tier Feature Distillation Multiple Instance Learning for Histopathology Whole Slide Image Classification”.

[3] Zheng et al., “Kernel Attention Transformer for Histopathology Whole Slide Image Analysis and Assistant Cancer Diagnosis”.

[4] Jafarinia et al., “Navigating the MIL Trade-Off: Flexible Pooling for Whole Slide Image Classification”.

[5] Chen et al., “Scaling Vision Transformers to Gigapixel Images via Hierarchical Self-Supervised Learning”.

Questions for authors
  1. Contribution relative to ViT³. What architectural or optimization changes remain specific to SlideTTT? Could you evaluate against a ViT³-based MIL model with matched features, positional encoding, classifier capacity, and tuning budget, so that the benefit of those changes can be determined?
  2. Validity of the adaptation ablation. How does patch information reach CLS in the no-update and local-only variants? If disabling global updates removes this information path, could you bring a static control that maintains aggregation and isolates the effect of adapting the fast weights?
  3. Relevant aggregation and representation alternatives. Could you offer matched comparisons with R2T-MIL, DTFD-MIL, KAT, PANTHER, and HIPT, or describe specific barriers to a fair comparison? Such methods target different aspects of the feature, morphology, and spatial-modeling motivation, so their inclusion would clarify what benefit specifically requires TTT. For compatible architectures, evaluation with and without the proposed TTT module would also strengthen the modularity claim.
  4. Simple pooling controls. How does SlideTTT compare against standalone mean, max, LSE, and MaxSoft pooling under the identical feature and split protocol? Providing predictive performance together with aggregation cost could determine whether adaptation offers a worthwhile improvement over simpler alternatives.
  5. Trustworthiness of the observed gains. Are internal splits patient-disjoint, and were architecture choices and decision thresholds chosen using internal validation only? Please report class counts and uncertainty for paired external performance differences using patients as the resampling unit where patients contribute multiple slides.
  6. Agreement between metrics. What external PR F1 and balanced-accuracy values are correct in Tables 7/12 and 8/11? Please resolve these discrepancies and clarify how the AUC and balanced-accuracy results qualify the accuracy-based conclusions, especially for MSI and the plug-in experiments.
  7. Reliance on pretrained representations. Do the improvements remain with another frozen pathology encoder and with an encoder which was not pretrained on pathology, such as an ImageNet-pretrained ResNet-50? Keeping features fixed across competing aggregators within each encoder setting would help determine whether the improvement is robust to the source and strength of pretraining.
  8. Further OOD evidence. Could you assess CAM16-to-CAM17 transfer, or the opposite, with aligned slide-level metastasis labels? Ideally, target centers should be absent from training. This would strengthen the breadth claim by testing transfer in a sparse-lesion setting rather than adding another within-cohort benchmark.
  9. Reproducibility of the update. What initialization and gradient normalization are used in Eqs. (5) and (8), and does outer training differentiate precisely through the inner updates or use an approximation? These choices determine the learned adaptation procedure and are essential to reproduce it.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.