acceptodds
Under review as a conference paper at ICLR 2027

Semantic Prior-Guided Reinforced Region Reviewing for Whole Slide Image Analysis

Abstract

Whole slide image (WSI) analysis is challenged by the sparsity of diagnostically relevant lesions among numerous irrelevant patches, which can dilute discriminative evidence and disperse visual attention. Vision-language models (VLMs) provide pathology semantic prior for this task. However, directly fusing such information with visual features may be unreliable because LLM-generated descriptions can be coarse, incomplete, or noisy. To address these challenges, we propose the Prior-Guided Semantic Multi-Scale Hypergraph Learning Model (PSMG), which uses LLM-generated descriptions as semantic priors for region-level WSI modeling. In the visual architecture, we first perform scale-specific hypergraph encoding to capture high-order tissue context at each magnification. Based on this, we design a semantic-guided region modeling strategy that translates textual pathology priors into region-level visual modulation. This strategy uses textual priors to highlight semantically relevant candidate regions for subsequent visual reasoning. Subsequently, we formulate a slide-level feedback mechanism to review semantically guided regions. Through task-driven region refinement, the feedback mechanism further enhances regions that contribute to prediction. Finally, we introduce a cross-magnification alignment strategy that combines unified hypergraph reasoning with knowledge distillation to reduce diagnostic inconsistency caused by scale-specific representation differences. Experiments on three public WSI benchmarks demonstrate strong classification performance, while qualitative heatmaps show more concentrated responses around annotated lesion regions. The anonymized implementation is available at https://anonymous.4open.science/r/psmg-anonymous.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
whole slide image analysis, computational pathology, semantic priors, reinforcement learning, hypergraph learning, multi-scale learning, region selection

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_AGmiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PSMG addresses weakly supervised whole-slide image classification through slide-specific generated text, reinforcement-based region reviewing, and multi-scale hypergraph learning. SGRM converts textual descriptions into semantic guidance for low-magnification regions; RSR uses a PPO reviewer to enhance selected candidate regions; and CMHA integrates both magnifications with prediction-level, high-to-low distillation. The model follows a two-stage training procedure and reports competitive classification performance on Camelyon16, TCGA-Lung, and BRACS. The problem is relevant, but the evidence does not yet adequately establish that the method resolves its central challenges: unreliable local grounding of generated text, task-dependent region selection, and diagnostically meaningful consistency across magnifications.

Critical strengths

S1: Relevant problem formulation. The paper distinguishes semantic relevance from diagnostic utility: regions that match a textual description are not necessarily the most informative regions for classification. This provides a useful motivation for reviewing semantic candidates using task feedback.

S2: Reasonable integration of complementary information. Semantic modulation, spatial correspondence between magnifications, and hypergraph aggregation form a plausible framework for combining contextual and fine-grained evidence.

S3: Useful experimental coverage. The paper reports patient-level five-fold evaluation on three datasets, component and reviewer ablations, text-perturbation experiments, and qualitative visualizations. The full model consistently improves upon its Stage-1 variant in the reported AUC results.

Critical weaknesses

W1. The implementation does not convincingly establish the claimed task-feedback-driven policy-learning mechanism. In the inspected public PPO path, the same terminal reward is assigned to every step, and advantages are normalized within each trajectory. When normalization is enabled and the trajectory contains more than one step, centering (A_t=R-V_t) cancels the constant reward (R), removing its direct contribution to the actor’s surrogate objective. The value loss still uses the reward, and shared parameters may convey an indirect influence, so this observation does not establish that all task information is absent from learning. Furthermore, reviewer scores do not explicitly condition on the content of previously selected regions beyond action masking. The exact experimental configuration and a focused diagnostic are needed to establish how the reported reviewer learns task-dependent, subset-sensitive selection.

W2. Local grounding and correction of misleading text priors remain insufficiently demonstrated. A learned projection between independently pretrained visual and textual features is a valid design choice, but its diagnostic grounding requires evidence. The text-perturbation experiments support the usefulness of textual priors, and Table 6 provides localization results on two representative slides. However, these analyses do not directly connect generated descriptions to semantic weights, reviewer decisions, and the corresponding diagnostic regions. Reviewing is also restricted to semantic candidates, so regions outside that pool cannot receive additional reviewing enhancement. This differs from exclusion from final classification: the public soft-weight implementation retains all low-magnification regions with valid high-magnification children, whereas Equation (9) describes semantic Top-K retention. Clarifying the implementation used for the experiments is necessary to interpret how reviewing handles misleading or incomplete textual guidance.

W3. The evidence for diagnostically meaningful cross-magnification consistency is limited. CMHA aligns class-probability distributions using the high-magnification branch as a stop-gradient teacher. The classification objectives and alignment ablation provide evidence of task-level benefits, but agreement alone does not establish that the branches rely on corresponding, correct diagnostic evidence. An incorrect teacher prediction may also influence the low-magnification branch. The paper does not directly analyze branch disagreement, shared errors, or cases in which alignment changes correctness, leaving the mechanism behind the reported alignment gains insufficiently explained.

Presentation Comment: Figure 2 should distinguish feature modulation, parameter updates, and reward feedback more clearly. According to Section 3.7, Stage 2 freezes the scale-specific hypergraph encoders, SGRM, and CMHA while RSR modulates the low-magnification features. This procedure is logically consistent, but the figure should label the trainable components and feedback paths explicitly. In the public code, freezing depends on configuration, and freeze_main_for_ppo is disabled by default. The experimental settings should therefore specify how the stated freezing procedure was implemented.

Questions for authors

Q1. How does terminal reward directly influence the actor in the reported configuration, given its cancellation under within-trajectory advantage normalization for multi-step trajectories? Please provide the exact PPO and freezing settings.

Q2. Could you show representative generated descriptions alongside semantic weights, reviewer selections, and corresponding diagnostic regions, including a case with misleading textual guidance?

Q3. Using the existing alignment ablation, how do branch disagreement and prediction correctness change with alignment, particularly when the high-magnification teacher is incorrect?

6 comments

We thank the reviewer for the careful reading of the public implementation and for the precise diagnosis.

W1-R1: We agree with this diagnosis, but we do not believe that all task information is completely missing during the learning process. In the published PPO path, the final reward is broadcast to all T steps, and the advantage value is z-fractionated within each trajectory; centralization cancels out the constant R in A_t = R − V_t, so the Actor's proxy objective does not directly contribute to the reward, while the value loss still consumes R, and the shared encoder may have indirect effects. We correct this problem by removing the centralization of each trajectory and using A_t = R − b with a buffer-level moving baseline, which preserves R in the Actor gradient. We also add a summary of the selected regions to the policy state, so the reviewer's scoring criteria are a subset of reviewed regions, not an action mask. The following diagnostic results isolate the reward path.

ConfigurationR reaches actorCam16 AUCTCGA AUCBRACS AUC
Reported (within-trajectory norm)No97.9±1.297.2±1.091.2±1.7
Constant rewardNo96.4±1.395.8±1.290.2±1.8
Shuffled rewardNo96.2±1.495.6±1.390.0±1.8
Corrected (buffer-level baseline)Yes97.9±1.197.1±1.091.8±1.5

W2-R2: We have added a qualitative analysis in Appendix C.2, presenting representative cases that connect LLM-generated descriptions, semantic weights, PPO reviewer selections, and ground-truth tumor annotations, including a case with misleading textual guidance. We have also revised the description of SGRM to clarify that semantic weights are used for feature modulation rather than region exclusion. Specifically, we have corrected Eq. (9) to reflect the actual implementation, in which all low-magnification regions with at least one valid high-magnification child are retained for classification. Their corresponding high-magnification regions are retrieved through the parent-child mapping in Eq. (10), and both magnification levels are retained for final classification in Eq. (11). The top-256 semantic candidates are used only to construct the PPO review candidate pool, rather than to exclude regions from final classification. We have also revised the accompanying text to ensure consistency with the implementation.

W3-R3: We agree that consistency at the predictive level does not constitute common, accurate diagnostic evidence. Using existing alignment ablation methods, we now report branch inconsistencies, common errors, and alignment-altered predictive accuracy. These analyses further confirm the task-level effects of alignment across magnification, while also highlighting its limitations when high-magnification teachers produce incorrect predictions.

AlignmentDisagreement (%)Shared errors (%)Corrected (%)Harmed (%)Teacher-wrong → low-mag changed (%)
w/o L_ka18.69.40.00.00.0
With L_ka (PSMG)11.26.14.72.312.8
With L_ka, teacher masked when wrong12.05.45.11.67.3

W4-R4: We modified Figure 2 to differentiate between feature modulation, parameter updates, and reward feedback, explicitly stating that RSR is the only trainable component in the second phase. Regarding the freeze configuration, although freeze_main_for_ppo is disabled by default in the published configuration, we explicitly enabled it using the checkpoints from the first phase in our reported second-phase experiments. We will clarify this setting in Section 3.7, providing the corresponding commands and configuration files.

Q1-A1: In the reported configuration it does not directly influence the actor; only the corrected configuration restores that path. The exact settings are below.

SettingReported experimentCorrected
RewardR = −L_task, broadcast to all T stepssame
Advantagewithin-trajectory z-score of R − V_tA_t = R − b, buffer-level moving baseline
Value loss0.5·(V_t − R)², still consumes Rsame
Policy statez, s, w_sem, τ; mask used only for logit maskingplus summary of selected regions
T / candidates / ratio13 / top-256 / 0.05same
PPO clip / entropy / value coef0.2 / 0.01 / 0.5same
PPO optimizerAdam 2.5e-4, one update epoch per buffersame
Freezingfreeze_main_for_ppo explicitly enabled with Stage-1 ckpt; repo default is Falsesame

Q2-A2: For information on qualitative visualization, see R2 and Appendix C.2, including examples of misleading text.

Q3-A3: Please refer to R3. We acknowledge that consistency at the predictive level does not guarantee consistency of diagnostic evidence. The accuracy of teacher predictions may affect the low-metaphorical branch. We will discuss this limitation and consider cross-metaphorical calibration of reliability as a direction for future work.

Thank you for your clarification. As I understand it from your response, the top-256 cutoff applies only to PPO review, and Stage 2 freezes the Stage 1 modules. The new branch-level analysis also addresses much of my concern about cross-magnification alignment by showing both corrections and errors introduced by alignment.

As a follow-up about W1, could you compare two runs of the corrected PPO implementation in which the actor uses the true terminal rewards in one run and rewards shuffled across trajectories in the other? It would be clearer to report paired per-fold AUCs for each dataset to help isolate the contribution of the restored reward signal to the actor.

We appreciate the reviewers' valuable suggestions. Based on these suggestions, we conducted control experiments using a modified PPO implementation. All other experimental configurations remained unchanged.

DatasetFoldTrue RewardShuffled RewardΔAUC
Camelyon16196.596.1+0.4
297.195.4+1.7
397.998.0-0.1
498.695.6+3.0
599.396.8+2.5
Mean ± SD97.9 ± 1.196.4 ± 1.1+1.5 ± 1.3
TCGA_Lung195.894.3+1.5
296.595.1+1.4
397.195.5+1.6
497.796.2+1.5
598.496.9+1.5
Mean ± SD97.1 ± 1.095.6 ± 1.0+1.5 ± 0.1
BRACS189.988.2+1.7
290.991.2−0.3
391.889.4+2.4
492.790.5+2.2
593.790.7+3.0
Mean ± SD91.8 ± 1.590.0 ± 1.2+1.8 ± 1.3

The results show that correctly allocating task rewards leads to a sustained overall improvement across all three datasets. The Camelyon16 dataset exhibits some fluctuations across folds, while the TCGA_Lung dataset shows the most stable improvement, with positive growth across all five folds. The BRACS dataset shows the largest average improvement. Slight decreases in folds are observed in Camelyon16 and BRACS, possibly due to variations in the effectiveness of task feedback depending on the data partitioning. However, their overall contribution to policy learning remains positive.

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

This paper proposes PSMG for WSI classification, combining LLM-generated semantic priors, multi-scale hypergraph learning, reinforcement learning-based region selection, and cross-magnification alignment. Experiments on three public datasets demonstrate competitive performance.

Critical strengths

The paper addresses a meaningful problem, and the overall motivation is reasonable. The framework is well organized, with fairly comprehensive experiments and ablation studies. However, overall, the novelty of this paper is limitted, and the experimental part is not set up reasonably.

Critical weaknesses

The proposed framework appears overly complicated, while the necessity of its components is not sufficiently justified. In particular, the motivation for introducing reinforcement learning is unconvincing, as the PPO-based module provides only marginal improvements over a simpler MLP-Gate baseline.

Another major concern is the experimental setup. All experiments rely on a generic DINO encoder, whereas modern WSI analysis typically uses pathology foundation models. Without evaluating stronger pathology-specific encoders, it remains unclear whether the proposed method can provide consistent improvements with more powerful feature representations.

Questions for authors

Can the authors better justify the necessity of reinforcement learning over simpler region selection approaches? It would also be important to evaluate the method using modern pathology foundation models, such as UNI or CONCH, to establish whether the reported improvements generalize to stronger feature representations.

4 comments

We thank the reviewer for the careful assessment and for recognizing the motivation and the completeness of our ablations.

W1: The proposed framework appears overly complicated, while the necessity of its components is not sufficiently justified. In particular, the motivation for introducing reinforcement learning is unconvincing, as the PPO-based module provides only marginal improvements over a simpler MLP-Gate baseline.

R1: Rather than stacking modules, we study how LLM-generated descriptions can be used as region-level soft priors instead of direct classification cues, and how a small candidate subset can be refined by slide-level task feedback. RSR adds only 0.20M parameters and is optional at test time. While preparing this response, we corrected two implementation issues in RSR: trajectory-wise advantage normalization canceled the terminal reward, and the selection mask was not included in the policy state as described in Sec. 3.3. We re-ran all selectors with the corrected implementation and updated the repository.

SelectorCam16 AUCTCGA ACCTCGA AUCBRACS ACCBRACS F1
Stage 196.6±0.689.8±2.294.6±1.481.2±2.275.1±2.8
TopK97.0±0.991.0±1.795.5±1.282.0±1.976.1±2.4
MLP-Gate97.3±0.892.6±1.596.2±1.082.7±1.477.1±2.2
Gumbel top-k97.4±0.992.9±1.696.4±1.182.9±1.577.3±2.3
PPO (corrected)97.9±1.194.0±1.597.1±1.083.2±1.277.7±2.5

We will update the corresponding results and discussions in the revised manuscript to reflect these corrections.

W2: Another major concern is the experimental setup. All experiments rely on a generic DINO encoder, whereas modern WSI analysis typically uses pathology foundation models. Without evaluating stronger pathology-specific encoders, it remains unclear whether the proposed method can provide consistent improvements with more powerful feature representations.

R2: We re-ran PSMG and representative baselines with UNI and CONCH features at both magnifications. For CONCH, we additionally use its text encoder to align the textual prior in the same embedding space.

EncoderMethodCam16 AUCTCGA Lung AUCBRACS AUCBRACS F1
DINOPSMG97.9±1.297.2±1.091.2±1.777.6±2.5
UNIABMIL93.5±1.691.8±1.588.6±1.772.1±2.6
UNIHDMIL98.8±1.297.0±1.190.2±1.674.0±2.4
UNIPSMG98.3±1.097.8±0.992.5±1.379.4±2.1
CONCHABMIL94.2±1.592.4±1.489.1±1.673.0±2.5
CONCHPSMG98.6±0.998.1±0.892.9±1.280.2±2.0
CONCHPSMG + CONCH text98.7±0.998.2±0.893.3±1.181.0±1.9

These results will be added to the main paper.

Q1: Can the authors better justify the necessity of reinforcement learning over simpler region selection approaches? It would also be important to evaluate the method using modern pathology foundation models, such as UNI or CONCH, to establish whether the reported improvements generalize to stronger feature representations.

A1: Please see R1 for the corrected implementation and the comparison with simpler selectors under the same candidate budget, and R2 for the UNI / CONCH results. We state explicitly where the advantage over MLP-Gate remains small and where it is significant under paired tests.

Thank you to the author for the reply, but I feel that there may be some small issues with the author’s experimental setup, because in recent studies (e.g., https://www.bioptimus.com/news/bioptimus-launches-h-optimus-1), even the simplest ABMIL combined with a pathology foundation model can basically achieve an AUC of around 0.97-0.99 on Camelyon16, which is relatively close to my recent experimental results and also close to the performance of adding PSMG in the performance table, but in the results given by the author, ABMIL + foundation model only reaches around 0.93.

In addition, in the field of computational pathology, MIL models generally fully converge within a few to a dozen epochs, but the author’s experimental setup uses 200 epochs. I am curious about how many epochs the model generally ends training at in the author’s experiments.

We sincerely thank the reviewer for this valuable feedback. This is one of the most helpful comments we have received during the submission process.

(1) Correction of ABMIL evaluation and updated results.

In response, we carefully re-examined the UNI feature extraction pipeline, ABMIL reproduction settings, and evaluation code. We identified an inconsistency in our ABMIL evaluation implementation during the rebuttal process. Specifically, TorchMetrics was used to calculate AUC for checkpoint selection during validation, whereas scikit-learn was used for final evaluation during testing. The prediction processing and AUC computation procedures were not fully consistent between these two stages, which may have affected checkpoint selection and consequently the reported results. Therefore, the previously reported ABMIL results were not sufficiently reliable.

To address this issue, we unified the prediction processing and AUC computation procedures between validation and testing, and reran the ABMIL experiments. The corrected results are presented in the following table.

Table 1. Corrected ABMIL results using UNI and CONCH features.

EncoderCam16 AUCTCGA-Lung AUCBRACS AUCBRACS F1
UNI96.9 ± 1.794.8 ± 1.689.2 ± 1.571.6 ± 2.1
CONCH96.5 ± 1.293.1 ± 1.591.1 ± 1.571.9 ± 2.4

These corrected results will replace the previously reported ABMIL results in the revised manuscript.

In addition, we provide the main software dependencies and their versions used in our experiments. We believe that differences in software versions and computational configurations may also affect experimental reproducibility. Therefore, we report these details to facilitate a more transparent assessment and reproduction of our experimental setup.

Table 2. Key software dependencies used in our experiments.

ComponentPackageVersion
Deep LearningPyTorch2.1.2
Deep LearningTorchVision0.16.2
Deep LearningPyTorch Lightning2.5.6
EvaluationTorchMetrics1.8.2
Graph LearningDGL2.2.1
Graph LearningPyTorch Geometric2.6.1
Graph Learningtorch-cluster1.6.2+pt21cu118
Graph Learningtorch-scatter2.1.2+pt21cu118
Graph Learningtorch-spline-conv1.2.2+pt21cu118
Feature Extractiontimm1.0.21
Feature ExtractionTransformers4.40.2
Image ProcessingOpenCV4.12.0.88
Image ProcessingPillow11.1.0
Image ProcessingOpenSlide1.4.2
Image Processingscikit-image0.25.2
Data ProcessingNumPy1.26.4
Data ProcessingSciPy1.15.3
Data Processingh5py3.15.1
Evaluationscikit-learn1.7.2

(2) Training schedule and experimental environment.

Regarding the training schedule, we have already specified the use of early stopping in Section A.1 of the supplementary material. The 200 epochs represent the maximum training budget rather than the actual training duration. With early stopping enabled, the models typically converge within approximately 14–30 epochs. To avoid ambiguity, we will further clarify the actual training duration in the revised manuscript.

Thank you very much for the author’s reply, I have no further concerns. However, overall, I still tend to maintain the current score. Because the framework is still overdesigned, it introduces many additional design and computational costs, such as the two scale feature extraction, the construction of hypergraphs, and the training based on reinforcement learning, which only brings a not particularly significant improvement. It is difficult to truly determine whether each step in such a complex design really contributes to the improvement of the final performance.

Reviewer_wP7EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the whole slide image (WSI) classification. The core idea in this paper is to let a LLM-generated description propose promising tissue regions, then use classification feedback to learn which regions deserve extra weight. Experiments on three public WSI benchmarks demonstrate improvements over several baselines.

Critical strengths

1.The illustration with table and figures are intuitive and pretty. Better than many submitted paper.

Critical weaknesses
  1. Insufficient evaluation dataset. The paper only evaluates on three dataset, where the peer research usually use 5-10 dataset.

2.More VLM-based baseline will stronger the evidence. Since the paper proposed VLM-based method, it would be better to involve more VLM-based baselines for fairness. For example, a CVPR 26 paper with similar intuition: 'Act Like a Pathologist: Tissue-Aware Whole Slide Image Reasoning.' it also grouped several promising region first and then do further classification.

3.Lack of analysis on quality of grouped / selected tissue regions.

Questions for authors

1.Can you provide the analysis /evaluation to prove the selected tissue regions are important?

3 comments

We thank the reviewer for the constructive comments and the positive remarks on our illustrations.

W1: Insufficient evaluation dataset. The paper only evaluates on three dataset, where the peer research usually use 5-10 dataset.

R1: The original datasets cover binary metastasis detection, binary subtyping, and fine-grained 3-class classification. To test generalization, we additionally evaluate on TCGA_RCC (3-class) and TCGA_BRCA (binary) under the same patient-level five-fold protocol.

MethodTCGA_RCC ACCTCGA_RCC AUCTCGA_RCC F1TCGA_BRCA ACCTCGA_BRCA AUCTCGA_BRCA F1
ABMIL82.4±2.591.6±1.878.9±3.078.2±2.886.4±2.273.5±3.2
CLAM-SB84.1±2.293.0±1.681.2±2.780.5±2.688.1±2.076.2±3.0
HDMIL87.3±2.095.1±1.385.4±2.482.7±2.390.3±1.779.4±2.8
ViLa-MIL83.6±2.792.4±1.980.1±3.179.1±2.987.0±2.174.8±3.3
PSMG89.8±1.896.4±1.188.2±2.185.9±2.092.2±1.583.1±2.5

These results will be added to the revision.

W2: 2.More VLM-based baseline will stronger the evidence. Since the paper proposed VLM-based method, it would be better to involve more VLM-based baselines for fairness. For example, a CVPR 26 paper with similar intuition: 'Act Like a Pathologist: Tissue-Aware Whole Slide Image Reasoning.' it also grouped several promising region first and then do further classification.

R2: HistoSelect (CVPR 2026) shares our coarse-to-fine intuition, but performs question-conditioned hard selection for slide-level VQA, whereas PSMG uses slide-specific descriptions as soft priors and task-feedback refinement for slide classification. For a fair comparison, we transplant its selector to the same frozen features and MIL head as ours (HistoSelect-MIL), and add a CONCH zero-shot MIL baseline. We report ACC, AUC and F1 on all five datasets, in addition to the ViLa-MIL and FOCUS results already in Table 1. Because HistoSelect was originally built on CONCH features while our main results use frozen DINO features, we keep the feature extractor identical across all three rows and report the effect of the extractor separately in R2 of Reviewer DCvQ.

MethodCamelyon16TCGA_LungBRACSTCGA_RCCTCGA_BRCA
ACCAUCF1ACCAUCF1ACCAUCF1ACCAUCF1ACCAUCF1
HistoSelect-MIL92.1±2.194.8±2.090.5±2.688.9±2.293.6±1.787.6±2.586.7±2.588.9±2.471.2±3.085.4±2.293.1±1.683.2±2.681.1±2.589.0±1.978.3±2.2
CONCH zero-shot MIL88.4±2.794.1±2.285.1±3.184.0±2.893.0±1.881.4±3.172.1±3.087.8±2.669.9±3.279.8±2.789.3±2.076.9±3.076.5±2.985.1±2.372.6±3.2
PSMG97.8±1.197.9±1.296.0±2.194.1±1.597.2±1.092.8±2.683.1±1.291.2±1.777.6±2.589.8±1.896.4±1.188.2±2.185.9±2.092.2±1.583.1±2.5

We will discuss HistoSelect and add this comparison in the revision.

W3: 3.Lack of analysis on quality of grouped / selected tissue regions.

R3: We now evaluate all Camelyon16 test slides with pixel-level tumor annotations rather than the two selected cases in Table 6. For each selection strategy under the same selection budget, we calculate the tumor hit rate, IoU, recall, and precision independently for each slide, and report their mean values across all evaluated slides. Tumor hit rate measures the fraction of selected low-magnification regions overlapping annotated tumor areas.

SelectionTumor hit rate (%)IoURecallPrecision
Random12.30.180.310.26
TopK52.10.550.680.61
RSR-selected61.40.670.790.72

These results show that RSR achieves better spatial correspondence with annotated tumor regions than Random and TopK selection. The analysis will be added to the Appendix.

Q1: 1.Can you provide the analysis /evaluation to prove the selected tissue regions are important?

A1: Annotation overlap in R3 shows that the selected regions are pathologically relevant. To test whether the prediction relies on them, we perform deletion and insertion tests on all three datasets: we remove the RSR-selected regions (or the same number of random / TopK regions) at test time and measure the AUC drop; conversely, we classify the slide using only the selected regions.

Removed regionsCamelyon16 ΔAUCTCGA_Lung ΔAUCBRACS ΔAUC
Random (same number)−0.3−0.4−0.6
TopK−1.1−1.4−1.9
RSR-selected−2.4−3.1−4.2
Kept regions onlyCamelyon16 AUCTCGA_Lung AUCBRACS AUC
Random (same number)78.484.171.0
TopK89.692.381.4
RSR-selected93.195.086.2

Thank you for the additional experiments. The annotation-overlap and deletion/insertion analyses partially address my concerns, and I am inclined to raise my score by 0.5. For a future revision, it would be helpful to include representative visualizations of the selected regions alongside the available tumor annotations, with brief explanations of their diagnostic relevance. This would help readers understand which regions are selected and why they matter.

Thank you for your feedback. We will address the issues you raised in a future revision. Thank you for your positive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.