acceptodds
Under review as a conference paper at ICLR 2027

SMART: Preserving Diagnostic Semantics for Unsupervised Domain Adaptation in Whole Slide Image Classification

Abstract

While pre-trained foundation models and Multiple Instance Learning (MIL) have advanced computational pathology, their clinical deployment remains severely hindered by cross-institutional domain shifts. Although Unsupervised Domain Adaptation (UDA) offers a practical solution to such discrepancies, directly applying general UDA paradigms to gigapixel Whole Slide Images (WSIs) often leads to negative transfer. This is because aggregated slide-level representations inherently entangle domain-specific confounders with task-relevant diagnostic semantics, exacerbating pseudo-label confirmation bias in self-training and marginalizing hard target cases under asymmetric domain shifts. To address these challenges, we propose Semantic Memory-Augmented Robust Transfer (SMART), a novel UDA framework that jointly optimizes across feature, label, and distribution dimensions. At the feature dimension, a convergence-aware variational information bottleneck dynamically purifies representations by filtering domain-specific confounders while preserving diagnostic semantics. Leveraging this purified latent space, we mitigate self-training confirmation bias at the label dimension by retrieving and incorporating reliable non-parametric source priors. Finally, at the distribution dimension, we address asymmetric domain shifts through an entropy-guided cross-difficulty mixup strategy. Extensive evaluations on real-world cohorts demonstrate consistent improvements over state-of-the-art methods. Designed as a plug-and-play framework, SMART can be readily integrated into existing bag-based MIL methods, substantially improving cross-institutional WSI classification. Code will be released.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Unsupervised Domain Adaptation, Whole Slide Images, Multiple Instance Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_vi97OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents SMART, which considers unsupervised domain adaptation for whole slide image classification through frozen pathology features and a trainable multiple instance learning pipeline. This consists of a residual variational information bottleneck that is aware of the convergence process, source-memory based retrieval of pseudo-labels, and reweighting of the target samples guided by entropy and latent mix-up. The EMA teacher provides representations and pseudo-labels.

Critical strengths
  1. Good amount of evaluation along with well-defined unsupervised procedure. There are multiple clinical tasks as well as cohort pairs in the paper, and a lot of information is included in the appendix that addresses the issues in evaluation. Improvements made have been quite significant.

  2. The appendix is not limited to removal of just one component. Table 10 presents static VIB vs. converging aware version comparison. Tables 11 and 12 evaluate retrieval relevance and pseudo label quality. Tables 13 and 15 differentiate between re-weighting, mixup, and difficulty aware pairing.

Critical weaknesses
  1. It is hard to disentangle the improvement from baseline stabilization. Appendix A.2 reveals that many baselines predict every target slide as LUSC in the leading transfer direction. The explanation and common initialization are useful additions, but they do not show competitiveness once we stabilize and tune the baseline without labels.

  2. The extra benefit of convergence awareness has not been fully disentangled. In Table 10, the static VIB in the remaining SMART framework achieves a relatively high AUC already, compared to an even higher AUC in CR-VIB. This is a good comparison, but it doesn't differentiate between loss-aware gating and scheduling regularization.

  3. The semantic-purification interpretation is not among the diagnostic measures. Retrieval purity, class separation, and AUC have good properties, but fail to single out domain confounder elimination. Residual pathway may retain the original confounded representation, while compression is not selective in removing the domain data. A class-conditional domain probing test or a similar controlled analysis will be required to differentiate between the new mechanism and standard regularization.

  4. The convergence gate and stochastic evaluation require a detailed description. Section 3.3 explains the transformation from EMA source loss to the convergence gate in qualitative terms, but fails to provide the full function, normalization, and update constants. The evaluation rule for the stochastic bottleneck also requires further elaboration.

Questions for authors

See weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer vi97 (Part 1/3)

W1: Disentangling SMART's Gains from Baseline Stabilization

We thank the reviewer for the insightful comments and valuable suggestions on clarifying the proposed mechanisms. To disentangle performance gains from baseline stabilization, we conducted additional experiments incorporating Static VIB, EMA teacher consistency, KL warm-up, and confidence-based pseudo-label supervision into DANN under the same unsupervised adaptation protocol and evaluation splits.

Specifically, EMA teacher consistency applies soft KL supervision to all unlabeled target samples. Linear KL warm-up additionally schedules the VIB regularization coefficient, while confidence-based self-training introduces hard pseudo-label CE for target samples with teacher confidence ≥0.70.

MethodCPTAC→TCGATCGA→CPTACTCGA→IPDIPD→TCGA
Original DANN51.41±3.1695.99±1.9181.16±2.7180.33±1.55
+ Static VIB + EMA Soft-KL Consistency91.48±2.1694.74±3.2183.36±2.1482.48±1.14
+ Linear KL Warm-up92.08±2.6688.56±7.3983.44±1.8182.49±0.85
+ Confidence-gated Hard Pseudo-label CE90.09±5.1189.57±5.9583.74±0.8781.08±0.69
SMART (Ours)95.58±0.7398.20±0.7786.10±1.9585.08±1.03

AUC (%). Each conventional variant builds upon the preceding configuration.

These results highlight two important findings. First, the severe single-class collapse observed in pathology UDA motivates our adoption of VIB-based regularization, which effectively mitigates such negative transfer. However, preventing collapse is only one aspect of robust cross-domain adaptation. Beyond feature stabilization, SMART further addresses pseudo-label confirmation bias and difficult target samples through SARA and REALM. Consequently, SMART consistently outperforms the strongest conventional baseline by 3.50, 2.21, 2.36, and 2.59 AUC points across the four directions.

Importantly, SMART's advantages extend beyond baseline stabilization and are not confined to the extreme-collapse setting. Our comprehensive evaluation spans nine adaptation directions across seven WSI cohort. SMART improves target-domain AUC over the corresponding source-only models in all nine directions and achieves the highest AUC among evaluated UDA baselines across all five original transfers. Furthermore, extensive experiments demonstrate its compatibility with four MIL backbones (DTFD-MIL, CLAM, TransMIL, and WiKG) and two pathology foundation models (CONCH and UNI2), with broad performance improvements across diverse architectures and feature representations (Appendices A.5 and A.9 of the original submission). Together with the consistently superior performance over strengthened DANN variants, these findings demonstrate that SMART's effectiveness cannot be explained solely by baseline stabilization, but reflects broadly transferable benefits from coordinated feature-, label-, and distribution-level adaptation.

AuthorsOpenReview ↗

Response to Reviewer vi97 (Part 2/3)

W2: Separating Loss-Aware Feedback and Joint Gating from Conventional Scheduling

To disentangle convergence awareness from conventional scheduling, we conducted a controlled two-factor ablation varying (i) the control signal (predefined linear/cosine schedules vs. source-loss-aware feedback) and (ii) the control scope (KL-only vs. joint regulation of residual preservation, Gaussian perturbation, and KL compression), while keeping all other SMART components and training settings unchanged.

MethodCPTAC→TCGATCGA→CPTACTCGA→IPDIPD→TCGAMean AUC
Linear KL-only93.57±2.7199.48±0.5083.93±3.5482.86±0.8189.96
Cosine KL-only94.78±1.7299.01±0.7382.36±3.9382.70±0.9489.71
Loss-aware KL-only92.91±1.0199.17±0.3684.10±1.7683.35±0.6989.88
Linear Joint Gating87.68±11.9198.41±1.4484.99±2.7982.26±1.1188.34
Cosine Joint Gating83.89±18.9698.40±1.1684.47±0.8782.79±0.8287.39
Loss-aware Joint Gating (SMART)95.58±0.7398.20±0.7786.10±1.9585.08±1.0391.24

AUC (%). Mean AUC is the average across four adaptation directions.

These results demonstrate the complementary benefits of loss-aware feedback and coordinated feature compression. Applying loss awareness to KL regularization alone does not reproduce the full model's performance, while replacing loss-aware feedback with fixed linear or cosine schedules for joint gating leads to lower performance, particularly under severe domain shifts. In contrast, SMART achieves the highest AUC in three of the four adaptation directions, including both brain tumor transfers.

More importantly, SMART achieves the highest four-direction mean AUC of 91.24%, outperforming all five alternatives by 1.28–3.85 percentage points. Specifically, it improves average AUC by 2.91 and 3.85 points over linear and cosine joint gating, respectively, and by 1.36 points over loss-aware KL-only regularization. Even compared with the strongest alternative, Linear KL-only (89.96%), SMART retains a 1.28-point average advantage.

These findings suggest that CR-VIB's effectiveness arises from coupling convergence-dependent feedback with coordinated regulation of residual feature preservation, stochastic perturbation, and KL compression, rather than merely scheduling the KL coefficient or applying either mechanism independently.

AuthorsOpenReview ↗

Response to Reviewer vi97 (Part 3/3)

W3: Testing Semantic Purification After Residual Fusion with Alignment and Probes

To directly examine the semantic-purification mechanism, we conducted additional class-conditional domain alignment and diagnostic-information probing analyses. We compare CR-VIB against Static VIB and Static VIB + Linear KL Warm-up to distinguish convergence-aware compression from conventional regularization, while keeping all other SMART components and training settings unchanged.

All analyses use the actual residual-fused slide-level representations, directly assessing domain alignment and diagnostic discrimination after residual fusion. Target labels are used exclusively for post-hoc evaluation, while diagnostic probes are trained using source labels only.

TransferMethodConditional MMD² ↓Diagnostic Probe AUC (%) ↑
CPTAC→TCGAStatic VIB0.471488.67
Static VIB + KL Warm-up0.437193.25
CR-VIB (Ours)0.285194.88
TCGA→IPDStatic VIB0.078779.55
Static VIB + KL Warm-up0.096280.70
CR-VIB (Ours)0.042882.44

Results are averaged over five splits. Conditional MMD² measures source–target distribution discrepancy within diagnostic classes; diagnostic probes are independently trained on source WSIs and evaluated on held-out target WSIs.

These results demonstrate two complementary effects. First, CR-VIB achieves substantially lower class-conditional domain discrepancy than Static VIB, with relative MMD² reductions of 39.52% and 45.62%, respectively. Second, it simultaneously improves diagnostic-probe AUC by 6.21 and 2.89 percentage points, respectively. CR-VIB also outperforms conventional KL warm-up on both measures, demonstrating a more favorable balance between cross-domain alignment and diagnostic discrimination.

Importantly, these improvements are observed in the final residual-fused representations, rather than only in the isolated bottleneck branch, indicating that the benefits of CR-VIB persist after residual fusion. Together, these findings provide complementary representation-level evidence that CR-VIB enhances class-conditional alignment while preserving transferable diagnostic semantics, supporting the semantic-purification mechanism proposed in Section 3.3 of the original submission.

W4: CR-VIB Gate Formulation, EMA Updates, and Deterministic Inference

We provide the explicit convergence-aware gating function and clarify the inference procedure.

Convergence-aware gate. As described in Section 3.3 of the original submission, LEMA(t−1)\mathcal L_{\mathrm{EMA}}^{(t-1)} denotes the EMA of the source-domain classification loss at the end of the preceding epoch. The EMA is updated at each training iteration with a momentum coefficient of 0.99 and initialized with the first observed value. The gate for epoch tt is then computed as

\max\left\lbrace \epsilon, \left[ \min\left( \frac{\kappa}{\mathcal L_{\mathrm{EMA}}^{(t-1)}},1 \right) \right]^{\gamma_g} \right\rbrace.$$ Here, $\kappa=5\times10^{-3}$ is the fixed scaling factor, $\epsilon=0.01$ is the lower bound, and $\gamma_g=3$ controls the transition sharpness. The same values were used across all tasks and splits. **Stochastic training and deterministic inference.** During training, CR-VIB follows Eq. (1) of the original submission: $$z_{\mathrm{vib},k}^d = \mu_k^d+g^{(t)}\sigma_k^d\odot\xi, \qquad \xi\sim\mathcal N(0,I),$$ $$z_k^d = (1-g^{(t)})v_k^d + g^{(t)}z_{\mathrm{vib},k}^d.$$ During evaluation, Gaussian sampling is disabled, and the corresponding mean representation is used: $$z_k^d = (1-g^{(t)})v_k^d + g^{(t)}\mu_k^d,$$ where $g^{(t)}$ is the gate associated with the evaluated checkpoint. No Monte Carlo sampling or averaging is performed.
Reviewer_CEWdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents an unsupervised domain adaptation (UDA) framework namely SMART for whole-slide image (WSI) classification. It uses pre-trained pathology foundation models and is based on multiple-instance learning (MIL). Their proposed framework is composed of three components, each designed for a particular purpose: CR-VIB as an information bottleneck to filter out domain-specific confounders; SARA to mitigate self-training confirmation bias; and REALM that uses entropy reweighting and mixup to handle difficult target samples. The authors evaluated their method on few medical benchmarks on lung, brain, and cervical cancer tasks, showing improvements especially on two large-scale public repositories, CPTAC-NSCLC to TCGA-NSCLC adaptation task.

Critical strengths
  • Cross-institutional variation is one of the main challenges in computational pathology, and this paper aims to tackle it, especially on the CPTAC-to-TCGA task where most standard domain adaptation methods could fail.

  • The authors provide ablation studies to isolate the effect of different components e.g., static VIB, as well as the patient-level evaluation splits make this a well-designed study.

Critical weaknesses
  • The novelty is somehow overstated. While SMART outperforms the baselines on CPTAC-to-TCGA, nine of those baselines actually collapsing (e.g. predicting a single class), Well, it’s easy to show large gains when the baselines completely collapse :)

  • More importantly, the authors' own ablations show that just adding a standard static VIB increases the AUC from 64.46% to 92.90%. It’s also not convincing to me that CR-VIB is doing something better than basic regularization tricks like KL annealing.

  • There is missing formulation of the convergence-aware gate, making it impossible to reproduce. In addition, the class-balancing formula (Eq 3.5) doesn't actually guarantee balanced classes, can authors comment on this?

  • The method also filters out uncertain target samples before optimization which contradicts the paper's claim that it actually helps the hardest cases.

Questions for authors
  • What is the exact formula for convergence-aware gate g(t)? And how does CR-VIB actually compare to a static VIB that simply uses standard KL warm-up?

  • Since standard UDA methods collapsed on the lung cancer task, can you compare your method against a more stable baseline, for example by using EMA self-training and static VIB?

  • Does REALM actually help difficult or minority samples? It would be beneficial to report the performance on the subset of data , e.g., the classification performance isolated just for the hard and easy or minority/majority subsets, especially for the highly imbalanced brain tumor dataset.

4 comments
AuthorsOpenReview ↗

Response to Reviewer CEWd (Part 1/3)

W1 & Q2: Comparing SMART with Stabilized DANN Using EMA and Static VIB

We appreciate the reviewer's constructive comments and valuable suggestions regarding the effectiveness of SMART. To disentangle performance gains from baseline stabilization, we conducted additional experiments incorporating Static VIB, EMA teacher consistency, KL warm-up, and confidence-based pseudo-label supervision into DANN under the same unsupervised adaptation protocol and evaluation splits.

Specifically, EMA teacher consistency applies soft KL supervision to all unlabeled target samples. Linear KL warm-up additionally schedules the VIB regularization coefficient, while confidence-based self-training introduces hard pseudo-label CE for target samples with teacher confidence ≥0.70.

MethodCPTAC→TCGATCGA→CPTACTCGA→IPDIPD→TCGA
Original DANN51.41±3.1695.99±1.9181.16±2.7180.33±1.55
+ Static VIB + EMA Soft-KL Consistency91.48±2.1694.74±3.2183.36±2.1482.48±1.14
+ Linear KL Warm-up92.08±2.6688.56±7.3983.44±1.8182.49±0.85
+ Confidence-gated Hard Pseudo-label CE90.09±5.1189.57±5.9583.74±0.8781.08±0.69
SMART (Ours)95.58±0.7398.20±0.7786.10±1.9585.08±1.03

AUC (%). Each conventional variant builds upon the preceding configuration.

These results highlight two important findings. First, the severe single-class collapse observed in pathology UDA motivates our adoption of VIB-based regularization, which effectively mitigates such negative transfer. However, preventing collapse is only one aspect of robust cross-domain adaptation. Beyond feature stabilization, SMART further addresses pseudo-label confirmation bias and difficult target samples through SARA and REALM. Consequently, SMART consistently outperforms the strongest conventional baseline by 3.50, 2.21, 2.36, and 2.59 AUC points across the four directions.

Importantly, SMART's effectiveness extends far beyond mitigating severe baseline collapse. Our comprehensive evaluation spans nine adaptation directions across seven WSI cohorts, with SMART improving target-domain AUC over the corresponding source-only models in all nine directions. Moreover, SMART achieves the highest AUC among evaluated UDA baselines across all five original transfer directions, while the additional EBRAINS experiments further demonstrate its robustness under diverse cross-institutional shifts (Appendix A.9 of the original submission). This generalizability also extends across model architectures and feature representations. Extensive experiments demonstrate SMART's compatibility with four MIL backbones (DTFD-MIL, CLAM, TransMIL, and WiKG) and two pathology foundation models (CONCH and UNI2), yielding broad performance improvements across diverse configurations (Appendices A.5 and A.9 of the original submission). Collectively, these results provide comprehensive evidence of SMART's general effectiveness and robustness across clinical tasks, institutional cohorts, and model architectures, demonstrating that its contributions extend well beyond stabilizing an isolated failure case.

AuthorsOpenReview ↗

Response to Reviewer CEWd (Part 2/3)

W2 & Q1: CR-VIB Gate Formulation and Comparison with KL Warm-Up

As described in Section 3.3 of the original submission, LEMA(t−1)\mathcal L_{\mathrm{EMA}}^{(t-1)} denotes the exponential moving average of the source-domain classification loss measured up to the preceding epoch. The convergence-aware gate used in our implementation is

g(t)=max⁡{ϵ,[min⁡(κLEMA(t−1),1)]γg}.g^{(t)}=\max\left\lbrace\epsilon,\left[\min\left(\frac{\kappa}{\mathcal L_{\mathrm{EMA}}^{(t-1)}},1\right)\right]^{\gamma_g}\right\rbrace.

Here, κ=5×10−3\kappa=5\times10^{-3} is the fixed scaling factor and ϵ=0.01\epsilon=0.01 is the lower bound, and γg=3\gamma_g=3 controls the transition sharpness. The same values were used across all tasks and splits.

Comparison with conventional KL scheduling. To examine whether CR-VIB's benefits can be reproduced by conventional KL scheduling, we conducted additional experiments replacing CR-VIB with Static VIB + Linear KL Warm-up, while keeping all other SMART components and training settings unchanged.

MethodCPTAC→TCGATCGA→CPTACTCGA→IPDIPD→TCGAMean AUC
Static VIB + KL Warm-up93.57±2.7199.48±0.5083.93±3.5482.86±0.8189.96
CR-VIB (SMART)95.58±0.7398.20±0.7786.10±1.9585.08±1.0391.24

AUC (%). Mean AUC is the average across four adaptation directions.

The results demonstrate the effectiveness of CR-VIB beyond conventional KL scheduling. Across four adaptation directions, CR-VIB achieves a higher mean AUC (91.24% vs. 89.96%, +1.28 percentage points), with improvements in three directions spanning both lung and brain tumor classification. Notably, both methods already achieve near-ceiling performance (>98% AUC) on the reverse NSCLC transfer. The overall advantage, particularly the consistent improvements across both brain tumor transfers, highlights the broader effectiveness of CR-VIB beyond mitigating severe prediction collapse.

Unlike conventional KL warm-up, which schedules only the regularization coefficient while retaining stochastic perturbations during early training, CR-VIB jointly regulates residual feature preservation, noise injection, and KL compression according to source-domain learning progress. This enables adaptive preservation of diagnostic semantics and compression as learning evolves. Together with the controlled Static VIB ablation in Table 10 of the original submission, these findings support the value of convergence-aware representation learning across heterogeneous domain shifts.

W3: Adaptive Class Balancing Versus Equal-Count Sampling for Minority Classes

Our class-balancing strategy is designed to mitigate majority-class dominance while preserving sufficient high-confidence target supervision, rather than enforcing rigid equal-count sampling. Specifically, the adaptive quota in Section 3.5 of the original submission reduces the dominance of abundant pseudo-classes while retaining informative candidates for REALM's class-aware cross-difficulty optimization.

To validate this design, we conducted controlled experiments on both directions of the imbalanced five-class brain tumor task, comparing No Balancing, Strict Equal-Count Balancing, and our Adaptive Balancing while keeping all other SMART components unchanged.

TransferSampling StrategyOverall Macro-F1Minority F1Majority F1
TCGA→IPDNo Balancing45.28±3.727.92±5.3270.19±2.91
Strict Equal-Count47.49±2.9213.58±5.0370.10±3.61
Adaptive (Ours)50.28±3.8419.10±7.9571.07±2.23
IPD→TCGANo Balancing46.67±2.745.36±5.5974.21±1.07
Strict Equal-Count45.14±1.110.51±1.1574.90±1.59
Adaptive (Ours)48.28±2.0111.12±3.8573.06±1.19

F1 (%). Minority and Majority F1 are macro-averaged over the two least-represented and remaining three diagnostic classes, respectively.

The results highlight the importance of balancing supervision reliability and coverage. Without balancing, abundant pseudo-class candidates can dominate self-training, limiting minority-class recognition. Conversely, strict equal-count sampling may discard substantial high-confidence supervision, restricting the diversity of samples available for class-aware optimization. Our adaptive strategy addresses both challenges by reducing majority-class dominance while retaining additional high-confidence candidates for cross-difficulty optimization.

Crucially, Adaptive Balancing achieves the highest minority-class F1 in both transfer directions, improving by 11.18/5.76 points over No Balancing and 5.52/10.61 points over Strict Equal-Count Balancing on TCGA→IPD and IPD→TCGA, respectively. These improvements contribute substantially to the higher overall Macro-F1 (50.28% and 48.28%), demonstrating that adaptive balancing improves the recognition of underrepresented diagnostic classes rather than merely adjusting sampling proportions.

AuthorsOpenReview ↗

Response to Reviewer CEWd (Part 3/3)

W4 & Q3: REALM's Benefits for Hard and Minority Cases Despite Confidence Filtering

SMART distinguishes pseudo-label reliability from sample difficulty. SARA-rectified confidence and class-aware quota selection provide reliable target supervision, while REALM uses the entropy of the original teacher predictions to emphasize challenging samples through Soft-DRO reweighting and cross-difficulty latent mixup. Importantly, confidence filtering controls pseudo-label supervision, while domain-adversarial alignment continues to utilize target representations. Thus, SMART combines reliable self-training with explicit difficulty-aware optimization rather than simply favoring easy samples.

To directly assess REALM's benefits, we conducted additional evaluations on fixed hard/easy subsets and minority/majority diagnostic classes across both brain tumor transfer directions. Hard/easy subsets were defined using the source-initialized teacher's prediction entropy before adaptation, ensuring identical test subsets across variants. Minority classes were defined as the two least-represented diagnostic categories in each target cohort. The w/o REALM comparison retains the same class-aware candidate selection and remaining SMART components, isolating REALM's additional optimization contribution.

TransferOverall Macro-F1Minority F1Majority F1Hard Macro-F1Easy Macro-F1
TCGA→IPD45.59→50.288.69→19.1070.19→71.0744.16→47.8641.02→48.85
IPD→TCGA44.94→48.280.54→11.1274.55→73.0643.46→45.8144.09→44.88

F1 (%), w/o REALM (with class balancing) → Full SMART. Minority F1 averages the two least-represented classes; Majority F1 averages the remaining three.

These results demonstrate two complementary benefits. First, REALM substantially enhances minority-class recognition beyond class balancing alone, increasing minority F1 by 10.41 and 10.58 percentage points. These improvements contribute 4.16 and 4.23 points to the overall Macro-F1 gains of 4.69 and 3.34 points, respectively, highlighting REALM's particular value for underrepresented diagnostic categories.

Second, REALM improves performance on initially high-entropy target slides, increasing hard-subset Macro-F1 by 3.70 and 2.35 points, while also improving easy-subset performance. Together with the improved hard-slide representation geometry reported in Table 14 of the original submission, these findings provide direct evidence that REALM strengthens difficult-case generalization and minority-class recognition beyond class-aware sample selection.

Thank you again for your constructive feedback. We have carefully addressed all the concerns and questions you raised in our rebuttal, supported by additional experiments and methodological clarifications. We would be happy to provide further clarification if any questions remain.

Reviewer_1hsUOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents SMART, a framework for unsupervised domain adaptation in whole-slide image classification. SMART combines three components: a convergence-aware residual variational information bottleneck (CR-VIB) to gradually suppress domain-specific information; source-assisted retrieval augmentation (SARA) to refine target pseudo-labels using diagnostically similar labelled source slides; and REALM, which couples entropy-guided Soft-DRO with cross-difficulty latent mixup to emphasise difficult target samples. The framework is evaluated on three clinical tasks. The original benchmark contains six cohorts and five directed transfers, while the appendix extends the brain-tumour study with EBRAINS to seven cohorts and nine transfers.

Critical strengths
  1. The empirical evaluation is broad for WSI domain adaptation. It covers multiple diseases, binary and five-class tasks, bidirectional transfers, an additional EBRAINS cohort, two pathology foundation models, and attention-, transformer-, and graph-based MIL backbones.
  2. SMART generally outperforms source-only and UDA baselines, with particularly large gains on the difficult CPTAC-NSCLC to TCGA-NSCLC transfer.
  3. The paper uses ablations, sensitivity analyses, limited-target-data experiments, retrieval and pseudo-label diagnostics, representation analyses, and computational comparisons to support the proposed design.
Critical weaknesses
  1. The proposed method combines a scheduled variational information bottleneck, source-memory-based pseudo-label refinement, entropy-based reweighting, and latent mixup. Although their integration is sensible, these ingredients are individually related to established techniques. The manuscript does not clearly demonstrate whether the gains arise from a genuinely new algorithmic principle, the specific interaction between components, or simply their combined use and tuning. Comparisons against the closest component-level alternatives—not only module-removal ablations—are needed to establish novelty.
  2. The paper shows results which are means and standard deviations over five target splits, but it is unclear whether independent optimisation seeds were also varied. No paired confidence intervals or hypothesis tests are reported.
Questions for authors
  1. Given that several baselines collapse to single-class prediction on CPTAC-NSCLC→TCGA-NSCLC, what source-validation-only hyperparameter tuning was conducted for each baseline, and were the tuning ranges and computational budgets comparable to those used for SMART?
  2. How would SMART determine convergence, select a checkpoint, or detect harmful adaptation for a new target institution without access to target labels, rather than relying on a fixed 100-epoch endpoint?
  3. Why are the complete UDA baseline comparisons not reported consistently for the cervical and EBRAINS experiments? Can the authors provide the same baselines, metrics, and class-wise analyses across all transfer settings?
6 comments
AuthorsOpenReview ↗

Response to Reviewer 1hsU (Part 1/4)

W1: Testing SMART's Novelty with Conventional Component Replacements

We thank the reviewer for the thoughtful feedback and constructive suggestions on strengthening our evaluation. To distinguish SMART's contributions from a simple combination of established techniques, we conducted additional component-level replacement experiments, retaining the overall framework and evaluation protocol across five splits:

  1. CR-VIB vs. Static VIB + KL Warm-up: We replace convergence-aware gating with conventional Static VIB and linear KL warm-up. Unlike CR-VIB, which jointly controls residual preservation, stochastic perturbation, and KL compression according to source-loss convergence, this alternative schedules only the KL regularization coefficient.
  2. SARA vs. Multi-prototype Retrieval: We replace SARA's complete source-slide memory with three cosine-clustered prototypes per source class, reconstructed each epoch using the current EMA teacher. The same Top-5 retrieval, similarity temperature, and teacher–retrieval probability fusion are preserved. This comparison evaluates the value of SARA's instance-level retrieval, which retains richer intra-class morphological diversity than compact class prototypes.
  3. REALM's Cross-Difficulty Pairing vs. Class-Constrained FIFO Mixup (CC-FIFO): We replace REALM's separate hard/easy memory banks and difficulty-aware partner selection with a shared 512-entry FIFO memory, randomly sampling partners from the same pseudo-class. Soft-DRO, pseudo-label supervision, and the mixup objective remain unchanged. This comparison evaluates explicit hard–easy complementary optimization against class-constrained memory-based interpolation.

We further evaluate a joint replacement variant, simultaneously substituting all three specialized designs with their conventional counterparts while retaining the remaining SMART framework and evaluation protocol. This directly tests whether combining these conventional alternatives can reproduce the benefits of the full model.

MethodCPTAC→TCGATCGA→CPTACTCGA→IPDIPD→TCGAAvg.
Best Prior Baseline (Table 1)57.68±1.2395.99±1.9182.19±2.8281.92±2.3179.45
CR-VIB → Static VIB + KL Warm-up93.57±2.7199.48±0.5083.93±3.5482.86±0.8189.96
SARA → Multi-prototype Retrieval92.67±4.8497.25±2.4885.16±2.0783.82±0.9789.73
Cross-Difficulty Pairing → CC-FIFO91.12±5.0698.58±1.2485.59±1.6684.75±1.0990.01
All Three Replacements (Joint)89.71±3.5894.37±2.2884.44±3.9481.55±2.2787.52
Full SMART95.58±0.7398.20±0.7786.10±1.9585.08±1.0391.24

AUC (%). Avg. is the unweighted mean across four transfer directions. The three individual replacements modify one specialized design at a time, while the joint variant applies all three replacements simultaneously. Best Prior Baseline denotes DTFD-MIL, DANN, CLAM, and CDAN, respectively, from Table 1 of the original submission.

These results provide two complementary findings. First, Full SMART achieves the highest four-transfer average AUC (91.24%), outperforming the three individual component replacements by 1.28, 1.52, and 1.23 percentage points, respectively. More importantly, Full SMART outperforms the joint conventional-replacement variant across all four transfer directions, achieving a +3.72-point average AUC gain, with individual gains ranging from +1.66 to +5.87 points. These results provide stronger evidence that the specific designs of CR-VIB, SARA, and difficulty-aware pairing offer benefits beyond their conventional counterparts.

Second, all three single-component replacement variants still outperform the strongest previously reported baseline in every transfer direction, demonstrating the effectiveness of SMART's overall adaptation framework even when an individual specialized mechanism is replaced. However, jointly replacing all three designs leads to a larger performance reduction, indicating that combining their conventional alternatives does not reproduce the full model's effectiveness. Together with the positive interaction between Soft-DRO and cross-difficulty mixup demonstrated in Table 13 of the original submission, these findings support both the individual algorithmic contributions and the value of their coordinated integration.

AuthorsOpenReview ↗

Response to Reviewer 1hsU (Part 2/4)

W2: Independent Training Seeds and Paired Statistical Significance

Our original evaluation employed five independently generated patient-level target splits with a fixed optimization seed, assessing robustness to target-data partitioning. To further evaluate optimization stability, we extended the experiments to three independent training seeds per split for both SMART and the strongest baseline in each transfer direction, yielding 15 training runs per method per direction. The strongest baselines (DTFD-MIL, DANN, CLAM, and CDAN) were selected based on the highest mean AUC in Table 1 of the original submission.

For statistical analysis, we first averaged the three seed results within each matched split, then calculated paired AUC differences, two-sided 95% confidence intervals, and paired t-tests across the five splits, applying Holm correction across the four transfer directions.

TransferBest BaselineBaseline AUCSMART AUCPaired ΔAUC (pp), 95% CIHolm-adjusted p
CPTAC→TCGADTFD-MIL57.89±1.2395.52±0.24+37.63 [35.99, 39.27]1.46×10⁻⁶
TCGA→CPTACDANN95.92±0.9398.40±0.45+2.48 [1.22, 3.75]0.0110
TCGA→IPDCLAM81.63±3.0385.89±1.23+4.26 [0.99, 7.53]0.0224
IPD→TCGACDAN81.17±1.3885.21±0.76+4.04 [2.32, 5.77]0.0086

AUC (%), reported as mean ± SD across five matched splits after averaging three training seeds per split. Confidence intervals are split-level paired t intervals (df=4); p-values are Holm-adjusted across four transfers.

SMART achieves statistically significant improvements across all four transfer directions after Holm correction (all adjusted p < 0.05), with every paired 95% confidence interval entirely above zero. Importantly, on the three transfers with already strong baseline performance, SMART delivers additional 2.48–4.26 percentage-point AUC improvements, demonstrating advantages beyond challenging adaptation scenarios.

Moreover, SMART exhibits low sensitivity to optimization randomness, with inter-seed standard deviations of the five-split averaged AUC ranging from only 0.12 to 0.27 percentage points. Together, these results provide strong statistical evidence for SMART's consistent performance advantages and robustness to both target-data partitioning and independent optimization seeds.

Q1: Baseline Hyperparameter Tuning, Search Budgets, and Fairness

We follow the official implementations and recommended hyperparameters of the original baseline methods, making only the necessary adaptations to WSI-level MIL. For compatible feature-level UDA methods, adaptation is performed on slide-level representations after MIL aggregation, consistent with SMART. All DTFD-based methods share the same source-pretrained checkpoint, selected using source-validation data, as well as identical data splits and frozen foundation-model features.

Hyperparameter selection in UDA is inherently challenging without labeled target validation data. Although source-validation performance can guide source-model selection, it does not necessarily reflect adaptation performance under target-domain shifts. We therefore adopt the established hyperparameter configurations of the original methods rather than conducting additional target-specific searches. These configurations demonstrate strong competitiveness in our WSI benchmark, with DANN achieving 95.99% AUC on TCGA→CPTAC and CDAN achieving 81.92% on IPD→TCGA.

SMART likewise employs a single globally fixed hyperparameter configuration across all adaptation directions, without cohort-specific tuning. Thus, neither SMART nor the baselines undergo additional per-target hyperparameter searches, ensuring a consistent label-free evaluation protocol. Further details on model selection and optimization are provided in Appendix A.3.1–A.3.2.

AuthorsOpenReview ↗

Response to Reviewer 1hsU (Part 3/4)

Q2: Label-Free Checkpoint Selection, Training Stability, and Negative Transfer

Checkpoint selection without target labels is a well-recognized challenge in UDA. We therefore adopt a predefined, globally fixed 100-epoch training protocol, ensuring reproducible evaluation without target-label leakage or cohort-specific checkpoint tuning. Staged warm-up, gradual target supervision, and learning-rate scheduling further support stable adaptation.

To evaluate checkpoint sensitivity, we examined all epochs from 80 to 100 across four transfer directions, comparing SMART against the highest-performing baseline in each direction from Table 1 of the original submission, rather than a single fixed competitor.

Epoch / MethodCPTAC→TCGATCGA→CPTACTCGA→IPDIPD→TCGA
Best Reported BaselineDTFD-MIL: 57.68±1.23DANN: 95.99±1.91CLAM: 82.19±2.82CDAN: 81.92±2.31
Epoch 8094.13±4.0999.01±0.7286.83±1.5585.25±1.00
Epoch 9094.94±1.9998.50±0.9286.43±1.5885.16±0.93
100 (Reported)95.58±0.7398.20±0.7786.10±1.9585.08±1.03
Oracle (Best of 100)96.29±0.9599.45±0.4587.34±1.5385.56±0.95

AUC (%). Baselines are the highest-mean-AUC methods from the original benchmark. Oracle denotes retrospective target-label-based checkpoint selection, used only for analysis.

The results demonstrate both strong late-stage stability and consistent competitiveness. Across all 21 checkpoints from epochs 80–100, SMART's mean AUC consistently exceeds the strongest reported baseline in every transfer direction. The mean AUC varies by at most 1.91, 1.33, 1.12, and 0.18 percentage points, respectively, throughout this interval. Moreover, the predefined epoch-100 results remain within only 0.48–1.25 AUC points of the retrospective oracle.

Prospectively detecting harmful adaptation without target labels remains challenging in UDA, as existing label-free model-selection criteria can be unreliable across different adaptation settings[1]. To further assess this risk under our fixed protocol, we conducted a retrospective paired comparison between SMART at the predefined epoch 100 and its corresponding source-only model across all nine adaptation directions and five patient-level splits per direction. Remarkably, SMART achieves positive target-domain AUC gains in all 45 matched split-level comparisons (0/45 negative-transfer cases), with even the worst-case split showing a +1.27 percentage-point improvement. Direction-wise mean AUC improvements range from +2.41 to +37.90 points.

These findings demonstrate that SMART maintains strong and stable adaptation performance across different domain shifts using a single predefined, label-free checkpoint protocol, without requiring target-specific early stopping or checkpoint optimization.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.