Superpixel Boundary Correction for Weakly-Supervised Semantic Segmentation on Histopathology Images

Hongyi Wu; Hong Zhang

REVIEW 4 major objections 6 minor 12 references

Superpixel Boundary Correction for Weakly-Supervised Semantic Segmentation on Histopathology Images

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Superpixel floodfill refinement raises weakly supervised breast tissue segmentation to 71.08% mIoU.

desk verdict Modest, plausible extension of WSSS-Tissue with superpixel floodfill; the mIoU gain is real but the boundary-correction claim is undercut by a regression on the LYM class. read the letter →

arxiv 2501.03891 v1 pith:G2EB4RJT submitted 2025-01-07 cs.CV

Hongyi Wu , Hong Zhang This is my paper · ORCID

classification cs.CV

keywords weaklysupervisedsemanticsegmentationhistopathologybreastcancerBCSSdatasetclassactivationmapsuperpixelfloodfilltumormicroenvironment

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper tries to show that a cheap post-processing step can solve the boundary problem of weakly supervised semantic segmentation on histopathology images. The authors take CAM pseudo-masks from a classifier trained only on image-level labels, merge CAMs from three network depths, and then cluster the raw image into SLIC superpixels and flood-fill each superpixel with its dominant predicted class. On the BCSS breast cancer dataset this produces a mean IoU of 71.08%, above the previous best of 70.55%, with the top per-class score on necrosis and no pretrained model or data augmentation. The authors' conclusion is that superpixel boundaries, because they follow natural color and texture edges in tissue, are a reliable correction signal for CAM masks in pathology.

What carries the argument

The central object is Algorithm 1, the superpixel floodfill refinement. SLIC (Simple Linear Iterative Clustering) partitions the H&E patch into superpixels using spatial distance and color intensity; then for each superpixel, the algorithm counts pseudo-mask labels within it and, if the dominant class covers more than a threshold tau of the superpixel's pixels, reassigns the whole superpixel to that class. The mask being corrected comes from multi-layer pseudo-mask fusion, where CAMs from three network depths are combined through weighted cross-entropy losses. The floodfill's work is to suppress unnatural outlier blocks in CAM-derived masks and push boundaries onto the color and texture edges that superpixels follow.

What would settle it

Remove Algorithm 1 from the pipeline and evaluate on the same BCSS test split; the central claim collapses if the mIoU drop is not reproducible or if sweeping S, m, and tau produces results below the reported 71.08%.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that the residual error in weakly supervised tissue segmentation is largely a boundary-scale problem, and that a multi-level superpixel correction is enough to fix it. By extracting CAMs at three different CNN depths and fusing them in the loss function, the method captures tissue boundaries at different granularities; then Algorithm 1 imposes the superpixel partition on the fused mask, wiping out outlier blocks and aligning segmentation edges with the visual tissue distribution. Experimental results report 71.08% mIoU on BCSS, which the paper presents as state of the art for weak supervision on that benchmark, ranking first on necrosis and second on tumor and stroma. The authors also report that refining the segmentation-stage output improves performance, while refining the earlier CAM-stage pseudo-mask does not, which they interpret as evidence that premature CAM modification discards useful image-level information.

Load-bearing premise

The load-bearing premise is that the small clusters drawn by the SLIC algorithm on H&E patches follow true tissue boundaries, so assigning each cluster the dominant CAM class improves the mask; if many clusters straddle two tissue types, or if the undisclosed parameters S, m, and tau are unsuitable, the floodfill would erase thin structures such as isolated lymphocytes.

Editorial extensions

If this is right

Weakly supervised segmentation can reach competitive accuracy on breast cancer tissue without pixel-level annotations, lowering the cost of tumor microenvironment analysis.
Because the reported gain is additive and requires no pretrained model or augmentation, the same floodfill correction can be dropped onto other CAM-based WSSS pipelines.
Refinement should be applied to the segmentation stage rather than to the raw CAM, since the paper finds first-stage refinement has no effect.
The per-class balance (first on necrosis, second on tumor and stroma) implies the method does not improve average IoU by sacrificing difficult classes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper, the dependence on SLIC parameters S, m, and tau means transfer to other organs (lung, colon, prostate) is not guaranteed until those parameters are reported and re-tuned for different tissue boundary statistics.
The threshold tau acts as an eraser for thin structures: with a high tau, isolated lymphocytes or narrow stromal strands could be swallowed by a neighboring dominant class, which may explain why LYM IoU (60.48%) trails the other classes.
A testable extension is to make the threshold class-aware or boundary-aware, so that superpixels intersected by known high-contrast edges are exempted from floodfill.
Because the method is pure post-processing, it could be composed with stronger pseudo-mask generators to push the boundary further.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

This is a simple, honest paper: take WSSS-Tissue's multi-layer CAM pseudo-masks, add a SLIC superpixel floodfill step, and report a new best mIoU on BCSS. I believe the numbers are what they say, but the boundary-correction story does not survive their own per-class table.

What's new is the specific combination and the benchmark result. The components are known—WSSS-Tissue supplies the multi-layer pseudo-supervision, and superpixel refinement of CAMs exists in the wider WSSS literature. Applying this to breast cancer histopathology and showing a small improvement is a legitimate extension, not a conceptual advance. To the paper's credit, the writing is clear, the setup is standard, the dataset is public, and the total mIoU is internally consistent with the per-class numbers.

The big soft spot is the LYM class. WSSS-Tissue got 0.6098 LYM IoU; after the floodfill it drops to 0.6048. That is exactly the signature of majority-vote smoothing: lymphocytic infiltrate is the most fragmented class, so within mixed-tissue superpixels the minority pixels get overwritten. The paper never isolates Algorithm 1 in an ablation, so we do not know whether the +1.38 mIoU over WSSS-Tissue comes from genuine boundary alignment or from a smoothing bias that helps the larger classes. The SLIC parameters (S, m, tau) are not reported. There are no error bars or significance tests, and the gain over PistoSeg is 0.53 points, which could easily be noise. The baselines are quoted from their papers rather than re-run.

None of this is fatal. The method does improve overall mIoU, and the per-class gains on TUM, STR, and NEC are consistent. The LYM drop is small. But it undercuts the 'significantly improving boundary delineation' claim, at least for the class where boundary precision matters most.

The paper is for researchers working on CAM post-processing or weakly supervised segmentation in pathology. It is a useful data point on a public benchmark, and the failure mode (erosion of small objects) is worth knowing about. I would not build a project on the SOTA claim, but I would cite it as related work.

A serious editor should send this to peer review. The result is specific and checkable, and the issues are addressable with a proper ablation, hyperparameter disclosure, and significance testing. With those, the claim might hold; without them, the contribution is a modest empirical trick.

Referee Report

4 major / 6 minor

Summary. The paper proposes a weakly supervised semantic segmentation (WSSS) method for histopathology images that combines multi-layer CAM pseudo-mask fusion with a superpixel-based floodfill refinement. Using only image-level labels, the method trains a classification network (ResNet38d) and a segmentation network, then refines the final segmentation by assigning each SLIC superpixel the dominant predicted class when the majority fraction exceeds a threshold tau. On the BCSS breast cancer dataset, the method reports an mIoU of 71.08%, which is higher than the previously reported PistoSeg (70.55%). The paper claims state-of-the-art results and improved tumor microenvironment boundary delineation.

Significance. If the result is robust, the contribution is a simple post-processing addition to existing CAM-based WSSS pipelines for histopathology, with potential practical value because it requires no additional annotation. The authors provide a clear algorithm description, per-class IoU numbers, and qualitative visualizations. However, the manuscript currently does not substantiate that the improvement stems from genuine boundary correction: the per-class LYM IoU decreases relative to WSSS-Tissue, the key hyperparameters are unreported, no ablation isolates the floodfill module, and the comparison to baselines is not shown to be statistically grounded. The central claim is therefore plausible but not yet demonstrated.

major comments (4)

[Section 4.3, Table 1] The per-class results contradict the claim that the superpixel floodfill improves boundary delineation: compared with WSSS-Tissue [11], LYM IoU drops from 0.6098 to 0.6048, while the fragmented nature of lymphocytic infiltrate makes it exactly the class most vulnerable to majority-vote floodfill (Algorithm 1). Since no ablation isolates Algorithm 1 from the multi-layer fusion and training changes, the +1.38 mIoU over WSSS-Tissue cannot be attributed to boundary correction. Please add an ablation that toggles the refinement module on the same backbone and reports per-class changes.
[Section 3.2, Algorithm 1, and Section 4.2] The method's key hyperparameters—lambda_1, lambda_2, lambda_3, SLIC average cluster size S, compactness m, and threshold tau—are never reported, and no code is provided. Without these values or a sensitivity analysis, the experiments are not reproducible and the reported 71.08% mIoU cannot be independently verified or compared against variants.
[Section 4.3, Table 1] The paper does not state whether the baseline numbers (HistoSegNet, SEAM, C-CAM, WSSS-Tissue, PistoSeg) were re-run under the same protocol or quoted from the original publications, and it provides no error bars or significance tests. Given that the improvement over PistoSeg is only 0.53 mIoU points, the state-of-the-art claim requires either re-run baselines with matched evaluation or a statistical analysis over multiple seeds or splits.
[Section 4.1] The dataset description incorrectly states that pathologists 'classify the images without providing pixel-level mask information,' even though the evaluation uses pixel-level ground truth on the BCSS test set. This inconsistency should be corrected, and details of the patch generation and train/validation/test split should be clarified.

minor comments (6)

[Section 2.2] The description of SEAM [9] does not match the cited reference; reference [9] is 'Weakly supervised semantic segmentation by pixel-to-prototype contrast' (CVPR 2022), not the SEAM paper. Please correct the citation.
[Section 1] The phrase 'With siperpixels' should read 'With superpixels', and 'multi-layer segmenation backbone' has a typo.
[Section 3.2, Algorithm 1] The pseudocode uses the misspelled variable names 'classprefined' and 'prefined'; these should be 'classrefined' and 'refined'.
[Section 4.2] The text says '10 epoches' and 'training rate'; these should be '10 epochs' and 'learning rate'.
[Abstract and Section 4.3] The phrase 'great performance' is vague; consider reporting concrete numbers or stating the relative improvement explicitly.
[Figure 4] The figure caption does not explain how the 'SP-boundary' panel is produced; please clarify what is shown in that panel.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the derivation chain is self-contained and evaluated on a held-out test set.

full rationale

The paper's pipeline is a standard WSSS construction: a CNN classifier produces CAM pseudo-masks, multiple feature-map depths are fused via a weighted loss, SLIC superpixels are used to flood-fill dominant classes in the pseudo-masks, and a segmentation network is trained on the refined pseudo-masks. Each component is defined independently of the final mIoU metric: CAM is derived from classification weights and feature maps, the superpixel floodfill uses only image color/texture and a class histogram within each superpixel, and the final segmentation is a separate network trained with cross-entropy loss. No parameter is fitted to the test-set ground truth; the BCSS test split is held out, and the reported mIoU is an external evaluation. There are no self-citations, no uniqueness theorems, and no ansatz that is justified solely by reference to the authors' own prior work. The hyperparameters (λ1, λ2, λ3, τ, SLIC S and m) are not reported, but tuning them on a validation set would be ordinary practice and is not evidence of a definitional collapse. The paper's own Table 1 shows that LYM IoU drops from 0.6098 (WSSS-Tissue) to 0.6048 after the proposed refinement, which undermines the abstract's claim of 'significantly improving tumor microenvironment boundary delineation' for that class; however, this is a correctness or empirical-robustness concern, not circularity. The derivation does not reduce to its inputs and the predictions are not fitted to the evaluation labels.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a small set of untested assumptions about superpixel-semantic alignment and on several unreported hyperparameters. No invented entities are introduced, but the evaluation depends on the favorable behavior of SLIC flood-fill across tissue boundaries.

free parameters (4)

lambda_1, lambda_2, lambda_3
Weights for multi-depth pseudo-mask losses in the loss function; values not reported, presumably tuned on validation.
tau
Dominance threshold in Algorithm 1; controls which superpixels are flood-filled, value not reported.
SLIC average cluster size S
Controls superpixel scale; not reported in the experiments.
SLIC compactness m
Controls superpixel boundary smoothness; not reported in the experiments.

assumptions (4)

domain assumption Superpixel clustering reflects natural boundaries of an image
Invoked in Section 3.2 to justify replacing CAM boundaries with superpixel majority labels; no evidence provided for H&E tissue.
ad hoc to paper If a superpixel has a dominant predicted class, all its pixels should be assigned that class
This is the paper's heuristic in Algorithm 1; it can erase fine structures when superpixels cross tissue boundaries.
domain assumption Multi-label image-level classification can provide sufficient supervision for tissue segmentation
Standard WSSS assumption inherited from prior work; not re-established here.
domain assumption ResNet38d feature maps at three depths provide complementary scale information
Borrowed from WSSS-Tissue; no analysis in this paper supports the specific choice of depths.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Superpixel Boundary Correction for Weakly-Supervised Semantic Segmentation on Histopathology Images." pith.science (2026). https://pith.science/paper/G2EB4RJT

@misc{pith2026250103891,
  author       = {Pith},
  title        = {Pith review of: Superpixel Boundary Correction for Weakly-Supervised Semantic Segmentation on Histopathology Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G2EB4RJT}},
  note         = {Machine review of arXiv:2501.03891}
}

read the original abstract

With the rapid advancement of deep learning, computational pathology has made significant progress in cancer diagnosis and subtyping. Tissue segmentation is a core challenge, essential for prognosis and treatment decisions. Weakly supervised semantic segmentation (WSSS) reduces the annotation requirement by using image-level labels instead of pixel-level ones. However, Class Activation Map (CAM)-based methods still suffer from low spatial resolution and unclear boundaries. To address these issues, we propose a multi-level superpixel correction algorithm that refines CAM boundaries using superpixel clustering and floodfill. Experimental results show that our method achieves great performance on breast cancer segmentation dataset with mIoU of 71.08%, significantly improving tumor microenvironment boundary delineation.

Figures

Figures reproduced from arXiv: 2501.03891 by the authors.

**Figure 2.** Illustration of superpixels on WSI We utilize Simple Linear Iterative Clustering (SLIC) algorithm to partition superpixels. We control the SLIC algorithm through three parameters: average cluster size S and compactness m. In SLIC, the distance D of pixel vector space is defined for clustering: D = s ∆x S 2 + ∆y S 2 + ∆I m 2 , where x and y represent the coordinate of a pixel point while I represents the col… view at source ↗

**Figure 3.** Illustration of patches and masks in BCSS dataset. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Visualization of superpixel floodfill refinement for BCSS dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 11 canonical work pages

[11]

Multi-layer pseudo-supervision for histopathology tissue semantic segmentation using patch-level classification labels

Chu Han, Jiatai Lin, Jinhai Mai, Yi Wang, Qingling Zhang, Bingchao Zhao, Xin Chen, Xipeng Pan, Zhenwei Shi, Zeyan Xu, et al. Multi-layer pseudo-supervision for histopathology tissue semantic segmentation using patch-level classification labels. Medical Image Analysis, 80:102487, 2022

work page 2022
[1]

Suggestive annotation: A deep active learning framework for biomedical image segmentation

Lin Yang, Yizhe Zhang, Jianxu Chen, Siyuan Zhang, and Danny Z Chen. Suggestive annotation: A deep active learning framework for biomedical image segmentation. In Medical Image Computing and Computer Assisted Intervention- MICCAI 2017: 20th International Conference, Quebec City, QC, Canada, September 11-13, 2017, Proceedings, Part III 20, pages 399–407. Sp...

work page 2017
[2]

Semi-supervised medical image classification with relation-driven self-ensembling model

Quande Liu, Lequan Yu, Luyang Luo, Qi Dou, and Pheng Ann Heng. Semi-supervised medical image classification with relation-driven self-ensembling model. IEEE Transactions on Medical Imaging, 39(11):3429–3440, 2020

work page 2020
[3]

A comprehensive analysis of weakly-supervised semantic segmentation in different image domains

Lyndon Chan, Mahdi S Hosseini, and Konstantinos N Plataniotis. A comprehensive analysis of weakly-supervised semantic segmentation in different image domains. International Journal of Computer Vision, 129(2):361–384, 2021

work page 2021
[4]

Wsss4luad: Grand challenge on weakly-supervised tissue semantic segmentation for lung adenocarcinoma

Chu Han, Xipeng Pan, Lixu Yan, Huan Lin, Bingbing Li, Su Yao, Shanshan Lv, Zhenwei Shi, Jinhai Mai, Jiatai Lin, et al. Wsss4luad: Grand challenge on weakly-supervised tissue semantic segmentation for lung adenocarcinoma. arXiv preprint arXiv:2204.06455, 2022

arXiv 2022
[5]

Learning deep features for discriminative localization

Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2921–2929, 2016

work page 2016
[6]

Deep learning in medical image analysis

Dinggang Shen, Guorong Wu, and Heung-Il Suk. Deep learning in medical image analysis. Annual Review of Biomedical Engineering, 19(1):221–248, 2017

work page 2017
[7]

Whole slide imaging: technology and applications

Matthew G Hanna, Anil Parwani, and Sahussapont Joseph Sirintrapun. Whole slide imaging: technology and applications. Advances in Anatomic Pathology, 27(4):251–259, 2020

work page 2020

Show all 12 references

[8]

Histoseg- net: Semantic segmentation of histological tissue type in whole slide images

Lyndon Chan, Mahdi S Hosseini, Corwyn Rowsell, Konstantinos N Plataniotis, and Savvas Damaskinos. Histoseg- net: Semantic segmentation of histological tissue type in whole slide images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10662–10671, 2019

2019
[9]

Weakly supervised semantic segmentation by pixel-to- prototype contrast

Ye Du, Zehua Fu, Qingjie Liu, and Yunhong Wang. Weakly supervised semantic segmentation by pixel-to- prototype contrast. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4320–4329, 2022. 6 PRIME AI paper

2022
[10]

C-cam: Causal cam for weakly supervised semantic segmentation on medical image

Zhang Chen, Zhiqiang Tian, Jihua Zhu, Ce Li, and Shaoyi Du. C-cam: Causal cam for weakly supervised semantic segmentation on medical image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11676–11685, 2022

2022
[12]

Weakly-supervised semantic segmentation for histopathology images based on dataset synthesis and feature consistency constraint

Zijie Fang, Yang Chen, Yifeng Wang, Zhi Wang, Xiangyang Ji, and Yongbing Zhang. Weakly-supervised semantic segmentation for histopathology images based on dataset synthesis and feature consistency constraint. In Proceedings of the AAAI Conference on Artificial Intelligence, vo...

2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [11]

Multi-layer pseudo-supervision for histopathology tissue semantic segmentation using patch-level classification labels

Chu Han, Jiatai Lin, Jinhai Mai, Yi Wang, Qingling Zhang, Bingchao Zhao, Xin Chen, Xipeng Pan, Zhenwei Shi, Zeyan Xu, et al. Multi-layer pseudo-supervision for histopathology tissue semantic segmentation using patch-level classification labels. Medical Image Analysis, 80:102487, 2022

work page 2022

[2] [1]

Suggestive annotation: A deep active learning framework for biomedical image segmentation

Lin Yang, Yizhe Zhang, Jianxu Chen, Siyuan Zhang, and Danny Z Chen. Suggestive annotation: A deep active learning framework for biomedical image segmentation. In Medical Image Computing and Computer Assisted Intervention- MICCAI 2017: 20th International Conference, Quebec City, QC, Canada, September 11-13, 2017, Proceedings, Part III 20, pages 399–407. Sp...

work page 2017

[3] [2]

Semi-supervised medical image classification with relation-driven self-ensembling model

Quande Liu, Lequan Yu, Luyang Luo, Qi Dou, and Pheng Ann Heng. Semi-supervised medical image classification with relation-driven self-ensembling model. IEEE Transactions on Medical Imaging, 39(11):3429–3440, 2020

work page 2020

[4] [3]

A comprehensive analysis of weakly-supervised semantic segmentation in different image domains

Lyndon Chan, Mahdi S Hosseini, and Konstantinos N Plataniotis. A comprehensive analysis of weakly-supervised semantic segmentation in different image domains. International Journal of Computer Vision, 129(2):361–384, 2021

work page 2021

[5] [4]

Wsss4luad: Grand challenge on weakly-supervised tissue semantic segmentation for lung adenocarcinoma

Chu Han, Xipeng Pan, Lixu Yan, Huan Lin, Bingbing Li, Su Yao, Shanshan Lv, Zhenwei Shi, Jinhai Mai, Jiatai Lin, et al. Wsss4luad: Grand challenge on weakly-supervised tissue semantic segmentation for lung adenocarcinoma. arXiv preprint arXiv:2204.06455, 2022

arXiv 2022

[6] [5]

Learning deep features for discriminative localization

Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2921–2929, 2016

work page 2016

[7] [6]

Deep learning in medical image analysis

Dinggang Shen, Guorong Wu, and Heung-Il Suk. Deep learning in medical image analysis. Annual Review of Biomedical Engineering, 19(1):221–248, 2017

work page 2017

[8] [7]

Whole slide imaging: technology and applications

Matthew G Hanna, Anil Parwani, and Sahussapont Joseph Sirintrapun. Whole slide imaging: technology and applications. Advances in Anatomic Pathology, 27(4):251–259, 2020

work page 2020