REVIEW 4 major objections 6 minor 29 references
CP-Dilatation: A Copy-and-Paste Augmentation Method for Preserving the Boundary Context Information of Histopathology Images
T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read Dilating the copied tumor mask preserves boundary context and improves histopathology segmentation.
desk verdict A sensible augmentation tweak with a real blending bug that undermines the boundary-context story; deserves peer review but not citation in its current form. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the dilated source mask $M^D_{src}$, produced by convolving the augmented source mask with a kernel $K$ (cross, rectangular, or open shapes). Copying $O_{src} = X^{aug}_{src} \odot M^D_{src}$ captures the malignancy and its surrounding context; pasting with Gaussian blur $B(M^D_{src})$ blends the boundary. This mask also drives the new-label synthesis rule $M_{new} = \min(\max(M^{aug}_{tar} \ominus M^D_{src}, 0) \oplus M^{aug}_{src}, 1)$, which removes the covered target region and re-adds only the true source malignancy.
What would settle it
Train the same models on a control augmentation that pastes a random same-sized tissue patch uncorrelated with the tumor boundary, keeping the tumor copy itself identical; if Dice and IoU match CP-Dilatation, the benefit is not boundary context.
Extended reading notes
Core claim
The paper's central claim is that applying a dilation operation to the binary mask in copy-and-paste augmentation preserves the boundary context information of the malignancy, and this preservation is what improves histopathology segmentation. Concretely, the dilated mask $M^D_{src} = M^{aug}_{src} \otimes K$ extracts both the tumor and its surrounding tissue from the source image; after Gaussian blurring the pasted region, the synthesized image retains the tumor while adding a contextual ring from another sample. The authors report that CP-Dilatation outperforms naive CP, CP-Simple, and TumorCP on both datasets and both segmentation architectures, and GradCAM visualizations show the classifier focusing closer to the ground-truth malignancy when trained with the dilated version.
Load-bearing premise
The surrounding tissue ring cut from another patient's image must still carry the same boundary information after being blended into a different tissue background; if the blend is biologically unrealistic or just adds noise, the gains could come from generic regularization rather than context.
Editorial extensions
If this is right
- Training segmentation models with CP-Dilatation yields consistent Dice, IoU, and pixel-accuracy gains over naive CP, CP-Simple, and TumorCP on DigestPath2019 and Warwick-QU for both U-Net and DeepLab v3+.
- The gains are achieved without extra annotations: the dilation reuses the available tumor mask and adds only a convolution and blur step.
- The benefit appears across two datasets with different magnification levels, suggesting the boundary-context advantage is not tied to one imaging scale.
- Providing a slightly enlarged support region around the malignancy helps the network concentrate on the ground-truth tumor, as shown by GradCAM.
- CP-Dilatation combines naturally with existing spatial and optical augmentations, since those are applied before the copy step.
Reading between the lines
- A testable implication the authors do not run: an ablation that pastes only the dilated ring (the contextual annulus without the tumor) onto targets. If Dice gains persist, the mechanism is boundary context; if they vanish, the gains likely come from the larger pasted region acting as regularization.
- The method can be viewed as a cheap label-margin augmentation; the same trick could plausibly transfer to other medical segmentation tasks with ambiguous boundaries, such as skin-lesion or lung-nodule segmentation, but this is an extrapolation.
- Because the paper reports averaged results over repetitions without confidence intervals, a multi-seed paired comparison would clarify whether the reported improvements are stable or within run-to-run noise.
- The similar performance across different kernel shapes suggests the exact geometry of the dilation is not critical, so an adaptive or learned kernel could be explored further.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes CP-Dilatation, a copy-and-paste augmentation method for histopathology image segmentation. The key idea is to dilate the source tumor mask before copying, so that the pasted region includes not only the annotated malignancy but also a surrounding context band, which the authors argue is diagnostically relevant because tumor boundaries in histopathology are often ill-defined. The source region is then blended onto the target image using a Gaussian-blurred version of the dilated mask, and a new mask is synthesized by subtracting the dilated area from the target mask and adding the original source mask. The method is evaluated on two public benchmarks, DigestPath2019 and Warwick-QU, with two segmentation backbones (U-Net and DeepLab v3+), and compared against several baselines including naive copy-paste, copy-paste with Gaussian blur, Simple Copy-Paste, and TumorCP. The paper reports consistent Dice, IoU, and pixel-accuracy improvements for the three chosen CP-Dilatation configurations, and includes a qualitative GradCAM analysis on a classifier.
Significance. If the reported improvements are real and are genuinely caused by the preservation of boundary context, CP-Dilatation is a simple, inexpensive augmentation that could benefit tumor segmentation in histopathology and other medical imaging domains with scarce annotations. The paper is clearly written and evaluated on two public benchmark datasets, and the proposed method is easy to implement. The qualitative GradCAM analysis is a useful attempt to interpret why the augmentation helps. However, the central attribution claim—that the gains come from boundary context rather than from a synthetic blending artifact—is currently undermined by an inconsistency in the blending equation, and the quantitative evidence lacks statistical support. These issues are addressable, but they need to be fixed before the paper's main claim can be accepted.
major comments (4)
- [Section 2.3, Eq. (3.1)] The blending operation in Eq. (3.1) is internally inconsistent. The target image is erased with the unblurred dilated mask MsrcD, while the source object is inserted with the blurred mask B(MsrcD). At the boundary of the dilated region, MsrcD is already 1 (so the target is fully removed), but B(MsrcD) is below 1 (so the source is only partially added). The result is a dark seam around the pasted tumor. A proper alpha blend would use B(MsrcD) in both terms, e.g., X_new = B(MsrcD) ⊙ O_src + (1 − B(MsrcD)) ⊙ X_tar. The text claims that the Gaussian blur 'mitigates heterogeneous information,' but the equation applies it only to the source term. This leaves open the possibility that the reported gains come from the network exploiting this synthetic edge artifact rather than from preserved boundary context. The authors should either replace Eq. (3.1) with a symmetric alpha blend or explicitly justify the asymmetric masking and demonstrate that the experimental conclusions are unaffected by the seam.
- [Section 3.2, Table 1] The reported performance improvements are small (e.g., Dice 0.687 vs. 0.672 on DigestPath2019 for U-Net, and 0.904 vs. 0.895 on Warwick-QU for DeepLab v3+) and are presented as single numbers without error bars, standard deviations, or significance tests. Section 3.1 states that 'average performance by conducting multiple repetitions of experiments' was reported, but neither the number of repetitions nor the variance is given anywhere. Without this information, the central claim that CP-Dilatation 'demonstrated superior performances compared to the comparative methodology' is not statistically supported. Please report the full distribution of results (e.g., mean ± std across at least five independent runs) and, if feasible, a paired significance test across repeated runs or bootstraps.
- [Section 3.1, experimental settings] All hyperparameters of CP-Dilatation (kernel type, kernel size, and Gaussian sigma) were selected based on the validation set, and Table 1 reports only the three best-performing configurations. While validation-based model selection is not circular, reporting only the best variants overstates the method's expected performance and makes the comparison with baselines less informative. The authors should clarify how many kernel/sigma configurations were tried, whether the selected configurations were stable across datasets and backbones, and ideally report the average performance over the full set of configurations tried or use a nested validation procedure.
- [Section 2.2, Eq. (2.1)] The definition of the dilated mask MsrcD is incomplete. Eq. (2.1) defines it as a convolution of the binary mask with a kernel K, but the text does not specify how the convolution output is converted into the mask used in Eqs. (2.2) and (3.1). If the output is thresholded or binarized, the threshold should be stated; if it is used with continuous values, the exact normalization should be described. This detail is necessary for reproducibility and for interpreting the blending behavior near the boundary.
minor comments (6)
- [Abstract] The abstract states that the proposed method is 'superior to the other state-of-the-art baselines chosen for comparison,' but the set of baselines is limited and 'state-of-the-art' is an overstatement given that several recent augmentation methods are not included; a more measured phrasing would be appropriate.
- [Throughout] There are typographical inconsistencies in the method name: 'CP-Dilatation' is used in the title and most of the text, but Section 3.1 refers to 'CP-Dilation,' and the running header reads 'CP-DILA TA TION.' Please unify the spelling.
- [Section 2.2, Figure 4] Figure 4 labels the kernels as being 'filled with values indicating how much to refer to the outside pixel,' but the depicted kernels appear to contain only binary 0/1 entries. If non-binary kernels are used, the exact values should be specified in the text or caption.
- [Section 1, contributions] The claim that this is 'the first CP methodology designed for histopathological images' is too strong given that the authors themselves cite TumorCP [18] for medical tumor segmentation and other CP variants exist in medical imaging. Please temper the novelty claim or provide a more careful literature comparison.
- [Section 3.3, GradCAM analysis] The GradCAM experiment is qualitative and uses a classifier trained only on DigestPath2019; it is not directly connected to the segmentation task or to the Dice/IoU improvements. A quantitative version (e.g., measuring attention overlap with the ground-truth boundary region) would strengthen the authors' interpretation.
- [General, reproducibility] No code or data-splitting scripts are provided. Given the small performance margins, releasing the exact preprocessing, augmentation probabilities, and training configuration would substantially help assess the reproducibility of the results.
Circularity Check
No significant circularity: CP-Dilatation is an empirical augmentation comparison whose claims rest on external benchmarks, not on its own definitions or self-citations.
full rationale
The paper proposes a data augmentation transform defined by Equations 2.1-3.2 and compares it on two public histopathology benchmarks. No prediction is derived from the method's own output, no parameter is fitted to the test set, and no load-bearing argument reduces to a self-citation. The hyperparameters (kernel type, kernel size, sigma) are explicitly selected on the validation set, which is standard model selection rather than test-set fitting. The performance claims are external empirical comparisons against CP-Naive, CP-Simple, and TumorCP, none of which are authored by the present authors. There is no claimed uniqueness theorem, no imported ansatz from prior self-cited work, and no known result being renamed as a new contribution. The only substantive weakness is a potential technical artifact: Equation 3.1 erases the target with the unblurred dilated mask while inserting the source with the blurred mask, which could create a seam around the pasted tumor. That is a correctness/validity concern about whether the reported gains reflect boundary context or a synthetic edge cue, not a circularity concern, because the measured performance is still an empirical outcome on external data rather than a quantity forced by the method's definition.
Assumptions & free parameters
free parameters (5)
- Dilation kernel type =
DILATE, RECT, OPEN
- Dilation kernel size =
10, 30, 40
- Gaussian blur sigma =
0.4, 0.7, 0.0
- Augmentation probabilities =
paug=0.33, CP probability=0.5, RandomResizePad probability=1
- RandomResizePad range =
(-0.9, 0.5)
assumptions (4)
- domain assumption Dilated source patch retains diagnostically meaningful boundary context after pasting and blurring onto the target image.
- domain assumption The two benchmark datasets are representative enough that gains transfer to other histopathology tasks.
- domain assumption Ground-truth masks are accurate enough for augmentation without introducing label noise.
- standard math Convolution and element-wise mask operations behave as described on binary masks.
Cite this review
Pith. "Pith review of CP-Dilatation: A Copy-and-Paste Augmentation Method for Preserving the Boundary Context Information of Histopathology Images." pith.science (2026). https://pith.science/paper/OF3TZ5YF
@misc{pith2026250704660,
author = {Pith},
title = {Pith review of: CP-Dilatation: A Copy-and-Paste Augmentation Method for Preserving the Boundary Context Information of Histopathology Images},
year = {2026},
howpublished = {\url{https://pith.science/paper/OF3TZ5YF}},
note = {Machine review of arXiv:2507.04660}
}
read the original abstract
Medical AI diagnosis including histopathology segmentation has derived benefits from the recent development of deep learning technology. However, deep learning itself requires a large amount of training data and the medical image segmentation masking, in particular, requires an extremely high cost due to the shortage of medical specialists. To mitigate this issue, we propose a new data augmentation method built upon the conventional Copy and Paste (CP) augmentation technique, called CP-Dilatation, and apply it to histopathology image segmentation. To the well-known traditional CP technique, the proposed method adds a dilation operation that can preserve the boundary context information of the malignancy, which is important in histopathological image diagnosis, as the boundary between the malignancy and its margin is mostly unclear and a significant context exists in the margin. In our experiments using histopathology benchmark datasets, the proposed method was found superior to the other state-of-the-art baselines chosen for comparison.
Reference graph
Works this paper leans on
-
[1]
INTRODUCTION With the rapid development of computing technology, deep neural networks (DNNs) are being applied to various practi- cal tasks across industries. Notably, deep learning algorithms grounded in computer vision are actively applied to digital pathology [1]. Histopathological image segmentation, in par- ticular, is widely recognized as a valuable...
work page Pith review arXiv 2025
-
[2]
We propose CP-Dilatation that is tailored to the specific characteristics of histopathological images. As far as we know, this is the first CP methodology designed for histopathological images and their malignancies
-
[3]
Our methodology demonstrated supe- rior performances compared to the comparative method- ology
The performance of our proposed method and a compara- tive methodology was evaluated objectively using publicly available datasets. Our methodology demonstrated supe- rior performances compared to the comparative method- ology
-
[4]
METHOD X aug src , Xaug tar , Maug src , Maug tar = Augspatial(Augoptical(Xsrc, Xtar), Msrc, Mtar) (1) MsrcD = M aug src ⊗ K (2.1) Osrc = X aug src ⊙ MsrcD (2.2) Xnew = {(1 − MsrcD ) ⊙ X aug tar } ⊕ {B(MsrcD ) ⊙ Osrc} (3.1) Mnew = min(max(M aug tar ⊖ MsrcD , 0) ⊕ M aug src , 1) (3.2) ImageMask Input Source Copy from Source Paste on 𝑋𝑡𝑎𝑟 𝑎𝑢𝑔, 𝑀𝑡𝑎𝑟 𝑎𝑢𝑔 𝑋𝑠𝑟𝑐...
-
[5]
At this time, to prevent this being2(i.e
We add the actual malignancy annotation in Osrc using ⊕ operation of M aug tar with the previous output. At this time, to prevent this being2(i.e. The malignancy in the source and tar- get is situated at overlapping positions.), we perform min(·) operation with 1. The result of these operations is a synthe- sized mask Mnew(Equation 3.2)
-
[6]
Experimental settings Implementation details We leveraged the Resnet34 [19] as the backbone network
EXPERIMENT 3.1. Experimental settings Implementation details We leveraged the Resnet34 [19] as the backbone network. The probability assignments for naive augmentation paug and CP-Dilatation were set at 0.33 and 0.5, respectively. However, a paug value of 1 was exclusively applied to RandomResizePad. Our optimization strategy in- volved using the Adam opt...
-
[7]
CONCLUSION We present an novel augmentation method that tackles the pervasive problems in the histopathology image segmenta- tion tasks: high labeling cost and the scarcity of training data. In this study, we have conceived that valuable contexts can ex- ist beyond malignancy, and thus devised a tailored dilatation operation to extract them. Our approach ...
-
[8]
Ramsha Baig, Maryam Bibi, Anmol Hamid, Sumaira Kausar, and Shahzad Khalid, “Deep learning ap- proaches towards skin lesion segmentation and classi- fication from dermoscopic images-a review,” Current Medical Imaging, vol. 16, no. 5, pp. 513–533, 2020
work page 2020
Show all 29 references
-
[9]
Breast cancer detection, seg- mentation and classification on histopathology images analysis: a systematic review,
R Krithiga and P Geetha, “Breast cancer detection, seg- mentation and classification on histopathology images analysis: a systematic review,” Archives of Computa- tional Methods in Engineering , vol. 28, pp. 2607–2619, 2021
2021
-
[10]
Application of deep learning in histopathology images of breast cancer: A review,
Yue Zhao, Jie Zhang, Dayu Hu, Hui Qu, Ye Tian, and Xiaoyu Cui, “Application of deep learning in histopathology images of breast cancer: A review,” Mi- cromachines, vol. 13, no. 12, pp. 2197, 2022
2022
-
[11]
Imagenet classification with deep convolutional neural networks,
Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hin- ton, “Imagenet classification with deep convolutional neural networks,” Communications of the ACM, vol. 60, no. 6, pp. 84–90, 2017
2017
-
[12]
Dataset growth in medical image analysis research,
Nahum Kiryati and Yuval Landau, “Dataset growth in medical image analysis research,” Journal of imaging , vol. 7, no. 8, pp. 155, 2021
2021
-
[13]
Data augmentation for low-resource neural machine translation,
Marzieh Fadaee, Arianna Bisazza, and Christof Monz, “Data augmentation for low-resource neural machine translation,” arXiv preprint arXiv:1705.00440, 2017
2017 arXiv
-
[14]
Adaptive data augmentation for image classification,
Alhussein Fawzi, Horst Samulowitz, Deepak Turaga, and Pascal Frossard, “Adaptive data augmentation for image classification,” in 2016 IEEE international con- ference on image processing (ICIP) . Ieee, 2016, pp. 3688–3692
2016
-
[15]
Instaboost: Boosting instance segmentation via probability map guided copy- pasting,
Hao-Shu Fang, Jianhua Sun, Runzhong Wang, Minghao Gou, Yong-Lu Li, and Cewu Lu, “Instaboost: Boosting instance segmentation via probability map guided copy- pasting,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 682–691
2019
-
[16]
Simple copy-paste is a strong data augmenta- tion method for instance segmentation,
Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph, “Simple copy-paste is a strong data augmenta- tion method for instance segmentation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognitio...
2021
-
[17]
A comparative study of cnn and fcn for histopathology whole slide image analysis,
Shujiao Sun, Bonan Jiang, Yushan Zheng, and Fengy- ing Xie, “A comparative study of cnn and fcn for histopathology whole slide image analysis,” in Inter- national Conference on Image and Graphics . Springer, 2019, pp. 558–567
2019
-
[18]
Colorectal carcinoma: Pathologic aspects,
Matthew Fleming, Sreelakshmi Ravula, Sergei F Tatishchev, and Hanlin L Wang, “Colorectal carcinoma: Pathologic aspects,” Journal of gastrointestinal oncol- ogy, vol. 3, no. 3, pp. 153, 2012
2012
-
[19]
Denise C DeCostanzo, Jules M Elias, and John C Chu- mas, “Necrosis in 84 ovarian carcinomas: a morpho- logic study of primary versus metastatic colonic car- cinoma with a selective immunohistochemical analysis of cytokeratin subtypes and carcinoembryonic antigen,” Internationa...
1997
-
[20]
Budding is useful to select high-risk pa- tients in stage ii well-differentiated or moderately differ- entiated colon adenocarcinoma,
Takashi Okuyama, Tetsuro Nakamura, and Masahiko Yamaguchi, “Budding is useful to select high-risk pa- tients in stage ii well-differentiated or moderately differ- entiated colon adenocarcinoma,” Diseases of the colon & rectum, vol. 46, no. 10, pp. 1400–1406, 2003
2003
-
[21]
Digestpath: A benchmark dataset with challenge review for the pathological de- tection and segmentation of digestive-system,
Qian Da, Xiaodi Huang, Zhongyu Li, Yanfei Zuo, Chen- bin Zhang, Jingxin Liu, Wen Chen, Jiahui Li, Dou Xu, Zhiqiang Hu, et al., “Digestpath: A benchmark dataset with challenge review for the pathological de- tection and segmentation of digestive-system,” Medical Image Analysis,...
2022
-
[22]
Gland segmentation in colon histology images: The glas challenge contest,
Korsuk Sirinukunwattana, Josien PW Pluim, Hao Chen, Xiaojuan Qi, Pheng-Ann Heng, Yun Bo Guo, Li Yang Wang, Bogdan J Matuszewski, Elia Bruni, Urko Sanchez, et al., “Gland segmentation in colon histology images: The glas challenge contest,” Medical image analysis, vol. 35, pp. 4...
2017
-
[23]
U-net: Convolutional networks for biomedical image segmentation,
Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-net: Convolutional networks for biomedical image segmentation,” in International Conference on Med- ical image computing and computer-assisted interven- tion. Springer, 2015, pp. 234–241
2015
-
[24]
Encoder-decoder with atrous separable convolution for semantic image segmentation,
Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam, “Encoder-decoder with atrous separable convolution for semantic image segmentation,” in Proceedings of the European confer- ence on computer vision (ECCV) , 2018, pp. 801–818
2018
-
[25]
Tumorcp: A simple but ef- fective object-level data augmentation for tumor seg- mentation,
Jiawei Yang, Yao Zhang, Yuan Liang, Yang Zhang, Lei He, and Zhiqiang He, “Tumorcp: A simple but ef- fective object-level data augmentation for tumor seg- mentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention . Springer, 2021, pp. 579–588
2021
-
[26]
Deep residual learning for image recognition,
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778
2016
-
[27]
Adam: A method for stochastic optimization,
Diederik P Kingma and Jimmy Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980, 2014
2014 arXiv
-
[28]
Microsoft coco: Common objects in context,
Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll ´ar, and C Lawrence Zitnick, “Microsoft coco: Common objects in context,” in European conference on computer vision. Springer, 2014, pp. 740–755
2014
-
[29]
Grad-cam: Visual explanations from deep net- works via gradient-based localization,
Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra, “Grad-cam: Visual explanations from deep net- works via gradient-based localization,” in Proceedings of the IEEE international conference on computer vi- sion, 2017, pp...
2017
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.