- 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 Currently, the reported baselines and the related work are outdated.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
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 graph | Accuracy (%) | AUROC (%) | Retention (%) | GFLOPs |
|---|---|---|---|---|
| 4-NN | 85.8 ± 1.8 | 94.1 ± 1.1 | 29 ± 3 | 16.8 |
| 8-NN (ours) | 88.4 ± 1.3 | 96.3 ± 0.7 | 28 ± 2 | 19.2 |
| 16-NN | 87.6 ± 1.5 | 95.7 ± 0.9 | 28 ± 2 | 23.4 |
| 32-NN | 86.8 ± 1.7 | 95.1 ± 1.1 | 28 ± 2 | 31.7 |
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 for a fixed neighborhood size, where 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).
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:
| Component | LENS | LENS+ |
|---|---|---|
| Graph layers | 1 GCN layer | 3 residual GCN layers |
| Pooling | Masked edge-mass softmax over active patches | [CLS] attention restricted to active patches |
| Design objective | Computational efficiency | Predictive accuracy |
| CPTAC-Lung accuracy | 88.4% | 91.2% |
| CPTAC-Lung GFLOPs | 19.2 | 24.1 |
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.
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 2has 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 2now also discussesMAMMOTH (ICLR 2026),Weers et al. (MIDL 2026), andSparseXMIL (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.