acceptodds
Under review as a conference paper at ICLR 2027

LENS: Learnable Edge Network Sparsification for Interpretable Histopathology

Abstract

Graph neural networks for whole-slide histopathology typically operate on dense spatial graphs that mix diagnostically relevant tissue interfaces with many uninformative connections, increasing compute and obscuring the tissue relationships used for prediction. We introduce LENS, a weakly supervised framework for task-driven per-slide topology learning: from slide-level labels alone, Graph neural networks for whole-slide histopathology typically operate on dense spatial graphs that mix diagnostically relevant tissue interfaces with many uninformative connections, increasing compute and obscuring the tissue relationships used for prediction. We introduce **LENS**, a weakly supervised framework for task-driven per-slide topology learning: from slide-level labels alone, **LENS** learns which tissue-patch edges to retain and freezes them at inference into a deterministic sparse adjacency. This sparse adjacency serves as both the computation substrate and a structural explanation, rather than a post-hoc overlay on a dense model. Across 5,530 slides spanning lung, kidney, and breast cancer using three encoders (**SimCLR**, **CTransPath**, **UNI2-h**), **LENS**/**LENS+** achieve the most favorable accuracy-sparsity-compute trade-off: they remove  72% of candidate edges while **LENS+** attains best or second-best accuracy on all four benchmarks, with up to 6.3x lower compute than the strongest dense graph baseline. Because **LENS** heatmaps are derived directly from the sparse graph used in the forward pass, they provide structural rather than post-hoc explanations: in a blinded, head-to-head pathologist evaluation against attention-based MIL and dense-graph attribution baselines, **LENS** achieves 97% RoI precision and 92% recall (F1 = 0.93), with at least 73% fewer false positives than these baselines, and 78% of retained edges are concentrated in tumor and tumor-stroma regions without any region-level supervision. Task-supervised topology learning offers a principled mechanism for jointly optimizing efficiency, accuracy, and interpretability in computational pathology. Code and models: https://github.com/LENS-network-ai/LENS.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Whole-Slide Image Classification, Graph Neural Networks, Graph Sparsification, Computational Pathology, Interpretable Machine Learning, L0 Regularization, Hard-Concrete Gates, Weakly Supervised Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_L6LxOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Learnable Edge Network Sparsification (LENS), which represents a patched whole-slide image as an 8-nearest-neighbor graph, where nodes are patches with UNI2-h derived features, and edges are learned with a structure learning mechanism. LENS and LENS+ outperform dense and sparse GNN baselines and MIL aggregation in three TCGA cohorts. The authors ablate with no edges, random edges, various patch encoders, and different number of neighbours in the constructed graphs. They also argue that LENS leads to more interpretable outcomes.

Critical strengths
  • The idea of modelling the aggregation of patch representations into a single slide representation through graph sparsification is interesting.
  • LENS (and sometimes LENS+) outperforms the reported baselines and the authors report paired t-tests.
Critical weaknesses

The reviewer is critically concerned with the following:

  • The deriving motivation behind LENS is unclear. If the main message is related to performance improvements in public benchmarks (TCGA), then the authors are encouraged to compare against the most recent pathology foundation models (PFMs), such as Virchow2 [1], H-optimus1 [2], etc. The models used for the encoder robustness ablation (CTransPath and SimCLR) do not completely represent the SOTA PFMs. If a performance improvement is observed across multiple SOTA PFMs with LENS, then, this would strengthen the paper. However, if the message is related to pooling and interpretability, then the authors are encouraged to compare against the most recent methods designed to augment the interpretability of multiple instance learning (MIL) algorithms, such as, but not limited to, xMIL [3] and MAMMOTH [4]. The work of [5] is worth considering too. Currently, the reported baselines and the related work are outdated.
  • Modeling a whole slide image as a nearest-neighbour patch graph is unconvincing in its current form. What is the underlying reason for this design choice? Moreover, given that there are various ways to construct a graph (other than kNN), such as Delaunay triangulation and thresholding, or other variants that the authors can find in [6] (please see Fig. 5), the choice of 8-NN patch graphs is undermotivated. Furthermore, the overhead of constructing a graph both during training and at inference is not adequately discussed, which is an important aspect in computational biology.
  • The introduction of two models (LENS and LENS+) is confusing since what differs between the two is underspecified.
  • Finally, LENS combines well-established existing components, and hence, lacks novelty.

[1] Zimmermann, Eric, et al. "Virchow2: Scaling self-supervised mixed magnification models in pathology." arXiv preprint. 2024.

[2] Bioptimus. H-optimus-1, 2025. https://huggingface.co/bioptimus/ H-optimus-1.

[3] Hense, Julius, et al. "xMIL: Insightful explanations for multiple instance learning in histopathology." NeurIPS. 2024.

[4] Shao, Daniel, et al. "Mixture of mini experts: Overcoming the linear layer bottleneck in multiple instance learning." ICLR. 2026.

[5] Jaume, Guillaume, et al. "Quantifying explainers of graph neural networks in computational pathology." CVPR. 2021.

[6] Brussee, Siemen, et al. "Graph neural networks in histopathology: Emerging trends and future directions." Medical Image Analysis. 2025.

Questions for authors

Please see the above weaknesses section.

(Minor) In the spatial perturbation ablation, do the authors train LENS from scratch with perturbed edges, or perturb the edges of the test set only?

8 comments
AuthorsOpenReview ↗

Response to Reviewer L6Lx (1/4): W1

We thank the reviewer for the detailed review and constructive feedback. We have revised the manuscript to address the concerns regarding motivation and recent baselines, the choice of candidate graph and its overhead, the distinction between LENS and LENS+, and novelty.


Motivation.

The contribution of LENS is neither improved feature extraction nor a better post-hoc explainer. Instead, LENS learns, using slide-level supervision alone, the sparse tissue-graph topology on which prediction is performed. This learned structure is intended to improve predictive performance, computational efficiency, and structural interpretability within a single framework. The sparse graph is learned during training and used directly for prediction. Its retained connectivity therefore provides a faithful account of which patch-to-patch interactions the model uses, rather than a post-hoc approximation of those interactions.


Encoders.

UNI2-h is a pathology foundation model. Our encoder ablation spans three different feature-learning approaches: lung-pretrained SimCLR, the pathology-specific transformer CTransPath, and the foundation model UNI2-h. LENS+ ranks first or second under all three encoders, while retained-edge density remains approximately 28–29% (Appendix I.1). Because LENS operates on precomputed patch embeddings, its graph-learning mechanism is compatible in principle with other pathology encoders, including Virchow2 and H-optimus-1. The consistency of the results across the three evaluated encoders supports the contribution of learned topology beyond the choice of feature extractor.


Interpretability-related work.

Section 2 has been expanded to discuss the three cited works and clarify their relationship to LENS:

  • xMIL (NeurIPS 2024) derives explanations of MIL predictions through layer-wise relevance propagation. LENS instead exposes the sparse connectivity used directly in its prediction computation, without requiring a separate post-hoc graph explainer.
  • MAMMOTH (ICLR 2026) improves task-specific patch representations through a parameter-efficient mixture of low-rank experts. Its primary contribution concerns feature transformation within MIL, rather than learning sparse tissue connectivity; the two approaches address complementary components of a slide-level prediction model.
  • Jaume et al. (CVPR 2021) investigate the quantitative evaluation of GNN explanations in computational pathology. Their work motivates the importance of assessing explanations against expert evidence. LENS similarly includes a pathologist-grounded evaluation, while studying the interpretability of a learned sparse prediction graph.

Recent baselines.

The revised manuscript adds a direct comparison with WiKG (CVPR 2024), a task-driven dynamic WSI graph-learning method, under matched features, patient-level splits, and model-selection procedures. On CPTAC-Lung, LENS+ achieves 91.2% accuracy and 97.7% AUROC, compared with WiKG's 90.8% and 96.9%, while requiring approximately 26% fewer GFLOPs (Table 1).

Section 2 now also discusses MAMMOTH (ICLR 2026), Weers et al. (MIDL 2026), and SparseXMIL (Medical Image Analysis, 2026). SparseXMIL uses sparse convolutions to model spatial context efficiently, whereas LENS learns which candidate tissue-graph edges to retain from slide-level supervision.

AuthorsOpenReview ↗

Response to Reviewer L6Lx (2/4): W2

W2. Why an 8-NN candidate graph, and its overhead

Choice of graph type.

The survey the reviewer cites, Brussee et al. (Medical Image Analysis, 2025; 204 GNN histopathology papers), already supports our design. In its Fig. 5 (Graph Type), patch graphs are the most common type, and they match the patch-level foundation-model features (UNI2-h) used by all methods in our comparison. The survey also notes (Sec. 2, footnote 2) that for tiled patches the region adjacency graph is equivalent to 8-NN with diagonal neighbors. Our 8-NN graph is therefore the standard patch-graph construction described there, not an arbitrary alternative to it. We compare it directly with 4-, 16-, and 32-NN (Appendix I.2, Table I.3). Cell graphs require nucleus segmentation at high magnification, and superpixel tissue graphs (e.g., Weers et al., MIDL 2026) change the node definition itself. Both are complementary to LENS, which learns which edges to keep on a given candidate graph and could in principle be applied to them. Finally, the same Fig. 5 shows that adaptive graph structure learning, the setting LENS addresses, appears in only a small minority of the surveyed works and is identified there as an emerging direction.


Why spatial k-NN.

Diagnostic tissue structures such as glands, tumor–stroma interfaces, and invasion fronts involve spatially neighboring regions. For interior patches on a complete regular grid, the eight nearest neighbors comprise four axis-aligned and four diagonal neighbors. This provides a natural local spatial neighborhood for tiled WSIs. Spatial neighborhood graphs are well established in WSI analysis, including Patch-GCN (Chen et al., MICCAI 2021).

Compared with Delaunay triangulation, 8-NN avoids an ambiguity inherent to regular square grids: the four corners of each grid cell are cocircular, allowing either diagonal in a valid Delaunay triangulation. Whereas Delaunay selects one diagonal per cell, 8-NN includes both, ensuring that all immediate spatial relationships are available for task-supervised selection. LENS can subsequently remove unnecessary edges, whereas edges excluded during candidate construction cannot be recovered. Tissue masking and slide boundaries can modify the available neighbors.


Why 8-NN.

The 8-NN graph defines the candidate edge set, rather than the final prediction graph. LENS learns which candidate edges to retain using slide-level supervision. The choice of candidate graph remains important because edges excluded from this set cannot subsequently be recovered by the scorer. We therefore evaluated its influence directly (Table I.3).

Candidate graphAccuracy (%)AUROC (%)Retention (%)GFLOPs
4-NN85.8 ± 1.894.1 ± 1.129 ± 316.8
8-NN (ours)88.4 ± 1.396.3 ± 0.728 ± 219.2
16-NN87.6 ± 1.595.7 ± 0.928 ± 223.4
32-NN86.8 ± 1.795.1 ± 1.128 ± 231.7

The 4-NN graph restricts the available spatial interactions and reduces accuracy by 2.6 percentage points relative to 8-NN. Increasing the neighborhood to 16-NN or 32-NN raises computational cost without improving accuracy or AUROC. The retained-edge sets for 8-NN and 16-NN exhibit a Jaccard similarity of 0.71, indicating substantial overlap in the selected connections. Together, these results identify 8-NN as the most favorable accuracy–computation trade-off among the evaluated candidate graphs.


Why learned selection rather than thresholding.

Under matched 28% edge retention and the same backbone, similarity-based thresholding underperforms LENS by 3.8 accuracy points, while random edge removal underperforms by 6.4 points (Table 4). GraphLSurv, an adaptive WSI graph sparsification method, is also included as a baseline in Table 1. These comparisons support the contribution of task-supervised edge selection beyond the choice of candidate graph and retention level.


Computational overhead.

The candidate graph is constructed from patch coordinates on a regular grid. Using coordinate lookup, its construction scales as O(n)O(n) for a fixed neighborhood size, where nn is the number of patches. The graph is constructed once per slide and cached for reuse, avoiding repeated neighborhood searches during training (Appendix E.2). The reported GFLOPs include the edge-scoring computation; the one-time construction of the cached candidate graph is excluded from the reported GFLOPs (Appendix E.3).

AuthorsOpenReview ↗

Response to Reviewer L6Lx (3/4): W3/W4

W3. LENS vs. LENS+

LENS and LENS+ share the same edge-scoring architecture, Hard-Concrete gating mechanism, and deterministic sparse-graph construction at inference. They differ primarily in message-passing depth and slide-level aggregation:

ComponentLENSLENS+
Graph layers1 GCN layer3 residual GCN layers
PoolingMasked edge-mass softmax over active patches[CLS] attention restricted to active patches
Design objectiveComputational efficiencyPredictive accuracy
CPTAC-Lung accuracy88.4%91.2%
CPTAC-Lung GFLOPs19.224.1

LENS prioritizes computational efficiency, achieving an accuracy-per-GFLOP ratio of 4.60. LENS+ improves accuracy by 2.8 percentage points at an additional cost of 4.9 GFLOPs. Its deeper residual GCN enables information propagation over multiple graph layers, while [CLS]-based aggregation integrates information globally across active patches. Both variants retain the benefits of task-supervised sparse connectivity. Section 3.4 summarizes this accuracy–efficiency trade-off, and Appendix D provides the full architectural specifications.


W4. Novelty

We agree that the Hard-Concrete relaxation is an established technique (Louizos et al., ICLR 2018), originally introduced for learning sparse neural-network parameters. Parameterized graph sparsification has also been investigated in prior work, including PTDNet and NeuralSparse, as discussed in Sections 1–2 and Appendix A.

To the best of our knowledge, prior work has not investigated this combination of Hard-Concrete edge sparsification, graph-normalized density control, and slide-level supervision for learning task-specific sparse topologies in computational pathology. This is consistent with the survey cited by the reviewer [6], where adaptive graph structure learning remains a small minority of GNN histopathology works (Fig. 5). A distinguishing challenge is that each WSI produces a graph with a different number of tissue patches and candidate edges, requiring the sparsity objective to regulate edge retention consistently across variable-size graphs.

The contribution of LENS lies in its formulation of task-supervised sparse topology learning for weakly supervised whole-slide image classification. Each WSI induces a distinct, variable-size tissue graph, and supervision is available only at the slide level. LENS addresses this setting through three integrated mechanisms:

  • Slide-specific edge selection: a relation-aware scorer conditions retention decisions on patch-pair features, allowing shared learned parameters to generate a different sparse topology for each WSI.
  • Graph-normalized sparsity control: the sparsity objective regulates the retained fraction relative to each slide's candidate-edge count, accommodating substantial variation in graph size. Empirically, realized retention remains at approximately 28 ± 2% across slide sizes and feature encoders (Table H.6, Appendix I.1).
  • Prediction-aligned structural interpretation: at inference, the learned gates produce a deterministic sparse graph that directly defines the model's message-passing connectivity. The retained edges therefore identify the interactions used in prediction, rather than being produced by a separate post-hoc graph explainer.

Experimental evidence. The controlled ablations isolate the importance of task-supervised edge selection. At matched edge retention and with the same backbone, random removal and similarity-based thresholding reduce accuracy by 6.4 and 3.8 percentage points, respectively, compared with LENS (Table 4). Removing all graph edges reduces accuracy by 7.0 points for LENS and 9.2 points for LENS+ (Table 5), demonstrating the importance of graph-mediated interactions. Together, these findings show that the improvement cannot be explained by sparsity alone.

LENS thus contributes a unified formulation in which slide-level supervision determines sparse tissue connectivity, graph-normalized regularization controls its density, and the resulting topology serves both prediction and structural interpretation.

Reviewer_pExSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a method which uses learnable edge sparsification as a weakly supervised framework for WSI histopatholgy classification with graph learning. It learns which connections between tissue patches to retain using slide-level labels, then uses the resulting sparse graph for classification. The retained edge weights also produce heatmaps that highlight relevant tissue regions. Experiments across datasets show competitive classification performance with about 70% of candidate edges removed. A pathologist study further shows that the heatmaps extracted align well with expert annotated regions.

Critical strengths

1.The proposed method has strong performance with a good balance between classification performance and computational cost on the Pareto frontiers. Learning which tissue connections to retain from only slide-level labels is well motivated, and the illustrated heatmap also provides a useful interpretation of the prediction.

2.The evaluation covers four datasets, providing evidence that the approach works across different tasks. The ablations further support the effectiveness of the learning process.

Critical weaknesses
  1. Relevance to the general AI community at ICLR is limited. I believe that the work at hand would be a better fit in a more applied conference or journal.

  2. The paper lacks direct comparisons with closely related learnable graph sparsification methods, such as PTDNet and NeuralSparse. Also, although LENS is evaluated on several datasets, the models are trained and tested separately within each dataset. It remains unclear how well the learned edge-selection model transfers to another cohort.

  3. There are many relevant graph learning for histopathology papers that should be mentioned and compared to. One relevant baseline is Weers et al 2026 which have developed a method with high fidelity explanations.

  4. The overall presentation needs some substantial refinement: figure 1 and C.1 have artifacts that are most likely stemming from large generative models, like misaligned lines and jagged edges; the fonts in the figures are generally too small; some of the notations and equations in the main body seem unnecessarily detailed, making them difficult to follow.

  5. The link in the abstract does not work: \url{https://github.com/LENS-network-ai/LENS}.

Weers, Alexander, et al. "From pixels to histopathology: A graph-based framework for interpretable whole slide image analysis." Medical Imaging with Deep Learning. 2025.

Questions for authors

1.Were there any experiments on whether this method has generalizability, in particular, have you tried to train a shared edge scorer with all datasets, how will the performance be affected?

2.How consistent are LENS's explanations across different training runs? For the same slide, do models with similar classification performance tend to highlight the same tissue regions?

3.Which edge mass pooling rule was used in the experiments: normalized edge mass, as described in Section 3.4, or softmax over edge mass, as specified in Appendix D and Algorithm L.4? If softmax was used, were patches with zero retained edge mass explicitly masked out? Otherwise they would still receive nonzero pooling weights, contradicting the claim that these patches cannot influence the prediction?

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

This paper proposes LENS, a weakly supervised framework for learning slide-specific sparse graphs for whole-slide histopathology analysis. A relation-aware edge scorer combines patch features with channel-weighted pairwise differences to parameterize Hard-Concrete gates over a spatial 8-NN candidate graph. Graph-normalized sparsity regularization and adaptive density control encourage consistent edge retention across slides of different sizes. At inference, the gates produce a deterministic sparse weighted graph that supports both prediction and topology-derived visualization. The framework has two variants, LENS and LENS+, which differ in graph-processing depth and pooling architecture. The authors evaluate classification performance on four public cohorts using three feature encoders, conduct extensive ablation studies, and assess heatmap alignment with expert annotations on ten slides. Additional experiments examine the applicability of the sparsification mechanism to survival prediction.

Critical strengths
  1. The experimental evaluation examines topology learning from several complementary perspectives. In particular, comparisons with random edge removal and similarity thresholding at matched retention provide useful evidence that edge selection matters beyond sparsity alone. Evaluations across multiple cohorts and feature encoders further support the applicability of the approach.

  2. The framework connects graph sparsification with an inspectable representation of the structure used for prediction. The blinded expert-RoI evaluation provides quantitative evidence of pathological relevance beyond qualitative heatmaps. Although limited to one expert and ten slides, this study offers a useful assessment of the learned structure and strengthens the practical motivation.

Critical weaknesses
  1. The baseline comparisons do not sufficiently establish the advantages of LENS over existing task-driven graph-learning methods for WSI analysis. WiKG [1], which learns dynamic patch relationships, is a particularly relevant comparison but is absent from the evaluation. The matched-retention controls demonstrate advantages over random and similarity-based edge selection, but they do not establish whether LENS provides a better accuracy–compute trade-off than an existing learned graph-construction method.

  2. Several inconsistencies in the method description affect the assessment of central claims about density control and structural faithfulness. The main text defines pooling weights proportional to retained edge mass, whereas the appendix and pseudocode specify softmax pooling, which requires an explicit mask to exclude inactive nodes. The expected-retention formulation also uses the annealed temperature in the main text but a fixed parameter in the loss pseudocode. These discrepancies leave it unclear whether the stated objective and explanation mechanism accurately describe the evaluated implementation. In addition, the expert-alignment procedure requires a more precise specification, and the correspondence between some visualizations and their accompanying descriptions needs clarification.

References

[1] Li, Jiawen, et al. “Dynamic Graph Representation with Knowledge-aware Attention for Histopathology Whole Slide Image Analysis.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 11323–11332.

Questions for authors
  1. Could the authors compare LENS and LENS+ with WiKG [1] using the same patch features, patient-level splits, and model-selection protocol, while preserving WiKG's original graph-construction mechanism? Reporting both predictive performance and computational cost, including edge scoring or graph construction for each method, would help establish the advantages over an existing task-driven WSI graph-learning approach.

  2. Could the authors clarify the pooling operation used in the implementation? Section 3.4 defines weights proportional to retained edge mass, whereas Appendix D and Algorithm L.4 specify softmax over edge mass. If softmax is used, are inactive nodes explicitly masked before normalization? How is pooling handled in the zero-edge ablation, where the stated active-node mask would exclude every node? These details are necessary to assess the claim that patches without retained edges cannot influence the prediction.

  3. Does the expected-ℓ0\ell_0 term use the annealed temperature τ\tau or the fixed parameter β=0.66\beta=0.66 specified in Algorithm L.3? Could the authors also explain how the optimized expected retention relates to the realized fraction of nonzero edges after deterministic inference and averaging the two directional gates? This clarification is important for understanding whether the reported density stability follows from the stated control mechanism.

References

[1] Li, Jiawen, et al. “Dynamic Graph Representation with Knowledge-aware Attention for Histopathology Whole Slide Image Analysis.” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 11323–11332.

3 comments
AuthorsOpenReview ↗

Follow-up to Reviewer bbVB: Complete cost accounting and W2

WiKG comparison and complete cost accounting (Q1)

Q1 asked for predictive performance together with computational cost, including edge scoring or graph construction for each method. We report that accounting in full below, with candidate-graph construction now included for LENS and LENS+ alongside the all-pairs construction already counted for WiKG.

Inference GFLOPs, CPTAC-Lung, construction included for every method:

MethodModel inferenceGraph constructionTotal
WiKG32.4included in model32.4
LENS19.20.6719.9
LENS+24.10.6724.8

The comparison is unchanged under this accounting: LENS+ exceeds WiKG on both accuracy (91.2 vs 90.8) and AUROC (97.7 vs 96.9) at lower total inference cost. The revised manuscript reports this accounting throughout, including Table 1 and the abstract.

Candidate-graph construction uses a KD-tree over patch coordinates, with construction-and-query complexity O(N log N), which supersedes the O(N) given in our earlier response. Standard FLOP counters trace tensor operations such as matrix multiplications and convolutions, and so do not capture the indexing, comparisons and tree traversal involved here; we therefore account for this step analytically. We charge the all-pairs cost of 6N² operations, the work a brute-force search would perform, as a conservative accounting proxy. Under this proxy, construction contributes 0.67 GFLOPs per slide, 3.4% of the LENS total. On the separate forty-slide timing set (TCGA-LUNG under our UNI2-h tiling), construction takes 15–16 ms per slide (median).

W2. Pooling, temperature, and expert-alignment protocol

The revised Section 3.4 states the masked softmax pooling used in the implementation together with the zero-edge fallback, consistent with Appendix D and Algorithm L.4; the expected-ℓ0\ell_0 term is stated with the fixed β=0.66\beta = 0.66 used in the code, with the annealed τ\tau confined to stochastic gate sampling; and Appendix J specifies the expert-alignment protocol in full, applying the same min–max and Otsu rule to all three methods with no threshold tuned against the expert annotations.

We would welcome knowing whether any part of W1 or W2 remains unresolved. If a specific element is still outstanding, we would be glad to address it within the discussion period.

The authors have addressed my concerns about the WiKG comparison, masked pooling, and the zero-edge fallback. The expert-alignment protocol is also more clearly specified. However, two numerical inconsistencies require clarification to establish the reliability of the reported alignment metrics.

  1. Precision and false-positive area in Figure J.1. The figure reports 98.1% precision and a false-positive area of 2.3% for LENS. Appendix J defines false-positive area as FP divided by total tissue area. Since the predicted mask is restricted to tissue, this quantity cannot exceed 1−precision=1.9%1-\mathrm{precision}=1.9\%. Please verify these values and their denominators.

  2. Case-level and aggregate recall for GTP. Figure J.1 states that GTP highlights no region and misses the entire expert RoI, implying zero recall for this LUAD case. Table J.1 reports a mean recall of 84.3% across five LUAD slides. If this case belongs to that evaluation, the highest possible mean recall would be 80%, even with perfect recall on the other four slides. Please clarify whether the figure and table use the same cases, model outputs, and binarization protocol.

Please provide the per-slide TP, FP, FN, and TN patch counts for each method on the ten evaluated slides, and reconcile the figures with the aggregate metrics. This would help verify the reported quantitative interpretability results.

AuthorsOpenReview ↗

Response to Reviewer bbVB

We thank the reviewer for these checks. We address both points below.

1. Precision and false-positive area

The implementation computes this quantity as FP divided by the non-RoI tissue area, FP/(FP+TN) — that is, the false-positive rate (FPR). The bound the reviewer applies, FP ≤ 1 − precision, follows from a total-tissue denominator, under which TP + FP ≤ tissue; with the negative class in the denominator that bound does not apply.

Section 4 of the main text refers to this quantity correctly as a false-positive rate. The Table 3 caption and the alignment-protocol paragraph in Appendix J describe the denominator as total tissue area; this is a terminological editing error, and the wording will be corrected to non-RoI tissue area in the revision:

  • Table 3 caption: "as a percentage of tissue area" → "as a percentage of non-RoI tissue area (false-positive rate)"
  • Appendix J, alignment protocol: the same substitution in the definition of FP
  • Column headers in Tables 3, 4 and J.1: "FP (%)" → "FPR (%)"

The correction affects the wording only, and recomputes no value in Table 3, Table 4 or Table J.1. All three methods are evaluated with the same denominator on the same tissue grid, so the comparison in Table 3 is unaffected: 2.9% for LENS against 10.8% for GTP and 12.1% for CLAM.

2. Recall for GTP on the LUAD case

The slide shown in Figure J.1 comes from an earlier annotation round and is not part of the ten-slide cohort used for the quantitative evaluation in Table 3 and Table J.1. This explains the discrepancy between the case-level visualization and the aggregate results: the five LUAD slides aggregated in Table J.1 do not include this case, and the behavior of any method on it does not enter the reported recall. We will replace Figure J.1 with a case drawn from the evaluation cohort, so that the figure and the tables derive from the same annotations, and we will report that case's precision, recall and false-positive rate in the caption. The caption's reference to the aggregate pattern will be revised accordingly.

We thank the reviewer for catching this. The aggregate metrics in Table 3 and Table J.1 are unaffected, as they are computed from the ten-slide evaluation cohort rather than from the figure. The illustrated case was selected for qualitative visualization independently of the ten-slide quantitative evaluation and was not used in computing the reported metrics for any of the methods. We will replace it with a case from the evaluation cohort to ensure consistency between the qualitative and quantitative evidence.

Per-slide counts

The per-slide TP, FP, FN, and TN counts require regenerating the binarized masks from the original evaluation. We prefer to provide these counts from a controlled, verified re-run rather than a hurried calculation against the discussion deadline. We will therefore reproduce the evaluation under the documented protocol and provide the complete per-slide counts for all three methods on the ten slides in Appendix J of the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.