Mitosis detection in domain shift scenarios: a Mamba-based approach

Francesco Tortorella; Gennaro Percannella; Mario Vento; Mattia Sarno

REVIEW 3 major objections 5 minor 14 references

Mitosis detection in domain shift scenarios: a Mamba-based approach

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A Mamba-based U-Net with stain augmentation generalizes better across unseen histopathology domains than a standard convolutional U-Net for mitosis detection.

desk verdict Competent and clearly written, but the leave-one-domain-out design leaks target-domain information into model selection, so the headline generalization claim outruns the evidence. read the letter →

arxiv 2508.21033 v2 pith:64MBSU6R submitted 2025-08-28 eess.IV cs.CV

Gennaro Percannella , Mattia Sarno , Francesco Tortorella , Mario Vento This is my paper

classification eess.IVcs.CV

keywords mitosisdetectiondomainshiftgeneralizationMambaVM-UNetstainaugmentationhistopathologyMIDOG

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 establish that replacing the convolutional blocks in a U-Net with Mamba state-space blocks improves mitosis detection when the model faces domains it never saw in training, and that random stain perturbation adds a further gain. On the MIDOG++ dataset, the authors run leave-one-domain-out experiments in which six domains train and the seventh is held out; their VM-UNet with stain augmentation reaches a mean F1 of 0.736, compared with 0.656 for a standard U-Net. The same approach scores 0.759 on the preliminary MIDOG25 track 1 test set. If the comparison holds, it gives histopathology practitioners a simple recipe—Mamba backbone plus stain augmentation—for making mitosis detectors less fragile across scanners, tissue types, and species.

What carries the argument

The central object is VM-UNet, a U-shaped encoder-decoder in which feature extraction uses Mamba VSS blocks (selective state-space models that scan image-patch sequences) instead of convolutional filters, together with patch embedding, patch merging/expanding, and skip connections. The second mechanism is stain augmentation: the Vahadane method estimates each image's stain matrix S and concentration matrix C, and training images are perturbed as I = I0 exp(−S(αC + β)) with random coefficients α and β, synthesizing new stain appearances. These two components carry the claimed generalization improvement.

What would settle it

Rerun the seven leave-one-domain-out folds selecting checkpoints by loss on a validation split carved from the six training domains, leaving the seventh domain untouched until final scoring; if the stain-augmented Mamba model no longer beats the U-Net by the reported margin, the generalization claim fails.

Watch

Extended reading notes

Core claim

The paper's central empirical claim is that a Mamba-based U-Net has higher generalization capability than a standard convolutional U-Net under domain shift, and that stain augmentation contributes positively on top of that. In the seven leave-one-domain-out experiments over MIDOG++ domains, the VM-UNet reaches a mean F1 of 0.710±0.073, versus 0.656±0.094 for U-Net; adding random stain perturbation raises the VM-UNet to 0.736±0.063. On the preliminary MIDOG25 track 1 test set, the approach obtains an F1 of 0.759. The authors interpret these results as evidence for the Mamba backbone and stain augmentation as complementary defenses against domain shift, while noting that the performance still

Load-bearing premise

The comparison only measures true generalization if the held-out domain is not used to choose the model; the paper selects the best model by loss on that held-out domain, so the reported advantage may be optimistic.

Editorial extensions

If this is right

On the MIDOG++ leave-one-domain-out protocol, switching from a convolutional U-Net to a Mamba-based U-Net raises mean F1 from 0.656 to 0.710.
Adding random stain perturbation to the Mamba model raises mean F1 further, to 0.736.
The approach reaches F1 0.759 on the MIDOG25 track 1 preliminary test set, showing the recipe is competitive in the challenge setting.
Tracing detection back to segmentation via NuClick masks lets the model be trained with dense Dice-plus-Focal supervision from point annotations.
The reported pipeline includes ensembling the best cross-validation models and morphological dilation of predicted masks to avoid splitting daughter cells.

Reading between the lines

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

The paper does not isolate whether the gain comes from Mamba's sequence modeling or simply from a different-capacity pretrained backbone; a matched-capacity convolutional U-Net with the same pretraining would separate these factors.
The stain-augmentation recipe is simple enough to transfer to other histopathology detection tasks where stain variation dominates, and the MIDOG++ leave-one-domain-out setup would be a ready benchmark.
A natural next step, not explored here, is combining the Mamba backbone with inference-time stain normalization or domain-adversarial training rather than augmentation alone.

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

Short version: this is a competent, incremental MIDOG25 challenge report whose central generalization claim is undercut by a model-selection protocol leak.

What is actually new is narrow: reported F1 numbers for VM-UNet, with and without stain augmentation, on the MIDOG++ leave-one-domain-out setup and a single test-set score. The components—VM-UNet, Vahadane-based stain perturbation, segmentation-then-detection with NuClick masks—are all published, and the combination is routine. But the paper is honest about being preliminary, and the experimental description is clear enough to reproduce. The workflow is sensible: balanced positive/negative batches, Dice+Focal loss, morphological post-processing, and ensembling of the best models.

The soft spot is the one the reader flagged, and it is load-bearing. The paper states that 'model selection was performed based on the value of the loss function on the validation domain.' In a leave-one-domain-out design, the validation domain is exactly the held-out domain. So the target domain participates in model selection before the reported F1 is computed on it. That is not a test of generalization to unseen domains. It can systematically favor whatever model overfits the validation domain best. The reported gaps (0.656 to 0.710 to 0.736) are plausible, but the standard deviations are 0.06–0.09 across only seven domains, and no significance test is reported. The sentence claiming higher generalization capability is stronger than the protocol supports.

The fix is straightforward: select models on training-domain data only, or justify why validation-domain loss selection does not inflate the result, and add per-domain results with significance tests. The ensembling of 'best found models' likely inherits the same leakage. Minor point: the abstract says 'large room for improvement'—that is fair, but the abstract also implies a robustness result that table 1 does not deliver.

This is a challenge report, not a methods paper. I would not cite the generalization claim until the selection is re-done, but the paper deserves a serious referee: the flaw is fixable, the benchmark is public, and the authors write cleanly. Send it to peer review if the venue accepts challenge papers; I'd ask for conditional acceptance pending the protocol fix.

Referee Report

3 major / 5 minor

Summary. The paper proposes a Mamba-based approach (VM-UNet) for mitosis detection under domain shift, framing detection as segmentation and adding stain augmentation to improve robustness. The method is evaluated on the MIDOG++ dataset using leave-one-domain-out experiments, comparing a standard U-Net, VM-UNet, and VM-UNet with stain augmentation. The authors report mean F1 scores of 0.656, 0.710, and 0.736, respectively, and an F1 of 0.759 on the MIDOG25 track 1 preliminary test set. The central claim is that the Mamba-based U-Net, especially with stain augmentation, has higher generalization capability on domains unseen at training time.

Significance. If the claim were supported, the paper would offer a useful data point on using state-space models for mitosis detection in histopathology under domain shift, a clinically relevant task. The MIDOG++ benchmark is public and the comparison is straightforward. However, the current experimental protocol has a critical flaw: model selection is performed on the held-out validation domain, which is the same domain used for final F1 evaluation. This makes the reported 'unseen domain' numbers a measure of transductive model selection rather than of generalization. In addition, the reported differences have overlapping standard deviations and are not accompanied by significance tests. The paper is short and preliminary, and these issues are load-bearing for its central claim.

major comments (3)

[Material and Methods, Experimental protocol] The leave-one-domain-out protocol uses the held-out domain as the validation domain: 'Model selection was performed based on the value of the loss function on the validation domain' and the reported F1 scores are computed 'on the validation domains.' This means the target domain is used during development to select the model (and, according to the Inference paragraph, to select the ensemble). Therefore, the results in Table 1 do not measure generalization to an unseen domain; they measure performance after model selection on the target domain, which can inflate F1 and bias model comparisons. The central claim that VM-UNet has higher generalization capability on domains unseen at training time is not supported by this protocol. Please use a protocol that keeps the target domain untouched until final evaluation (e.g., a validation split within the six training domains) or explicitly refram
[Results, Table 1] No statistical significance assessment is provided. With seven leave-one-domain-out folds, the standard deviations are large and the reported means overlap substantially: VM-UNet + Stain Aug (0.736 ± 0.063) differs from VM-UNet (0.710 ± 0.073) by only 0.026, and VM-UNet differs from U-Net (0.656 ± 0.094) by 0.054. Without paired significance tests across the seven folds (e.g., Wilcoxon signed-rank test) or confidence intervals, the statements that 'a Mamba-based U-Net has higher generalization capability' and that 'stain augmentation positively contributes' are not quantitatively supported.
[Material and Methods, Inference] The inference procedure states that 'model ensembling is exploited at inference time, by aggregating the predictions of the best found models in the performed cross-validation experiments.' It is not specified how the 'best found models' are identified, how many models are ensembled, or whether the selection uses the held-out validation domains. If the ensemble components are chosen based on the held-out domain, this is another form of target-domain leakage and makes the comparison against the single U-Net and single VM-UNet unfair. The ensemble composition and selection criterion must be stated.

minor comments (5)

[Abstract / Results] The abstract says 'Preliminary experiments ... show large room for improvement for the proposed method,' while the Results section states a stronger claim: 'a Mamba-based U-Net has higher generalization capability.' Please make the abstract consistent with the actual level of evidence.
[Material and Methods, Style augmentation] Equation (1) uses random coefficients α and β, but their ranges or sampling distributions are not given. For reproducibility, specify how α and β are drawn and whether the augmentation is applied only during training.
[Material and Methods, Model training] Training details are partial: no information is provided about the number of runs, random seeds, or variance across repeated training. Since the Table 1 statistics are over domains, not over training runs, it would be helpful to state whether each model was trained once or multiple times.
[Results, MIDOG25 test set] The F1 score of 0.759 on the MIDOG25 preliminary test set is reported without comparison to any baseline or description of the test set. As a single number it has limited interpretability; please clarify whether this is a challenge leaderboard result and how it relates to the leave-one-domain-out experiments.
[General] Minor language issues include 'significative performance drop' (Introduction), 'pertubation' (Abstract/Methods), and a reference formatting inconsistency in the bibliography. These do not affect the technical content.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: reported F1 scores are external benchmark measurements, not derived from fitted inputs or self-citations.

full rationale

This paper is a straightforward empirical evaluation. The derivation chain is: (i) the VM-UNet architecture is taken from prior external work (Ruan et al., VM-UNet) and pretrained on ImageNet; (ii) segmentation masks are generated from the dataset's point annotations using NuClick, an independent segmentation model; (iii) training uses a Dice+Focal loss on tiles, with stain augmentation via Vahadane stain separation; (iv) performance is measured on the held-out domain of the MIDOG++ leave-one-domain-out splits and on the MIDOG25 track 1 test set. None of these steps defines the reported F1 in terms of the model's own outputs or in terms of a parameter fitted to the evaluation set. The paper contains no self-citations: references [1]-[14] are to external method papers, datasets, and challenge rules. The central claim--VM-UNet plus stain augmentation generalizes better than U-Net--is supported by independently measured F1 scores on datasets not used for training. The protocol choice of selecting the best model by validation-domain loss is a test-set leakage concern that may weaken the generalization claim, but it is not circularity: it does not make the predicted quantity equal to an input by construction. Therefore the correct circularity finding is 'no significant circularity'.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new entities or mathematical machinery. Its claims rest on standard components (VM-UNet, NuClick, Vahadane stain estimation, Dice+Focal loss) plus two unreported tuning choices (augmentation coefficient ranges and ensemble recipe), and on three domain assumptions about tool reliability. The most consequential assumption is using the validation-domain loss for model selection, which biases the domain-generalization estimate.

free parameters (2)

Stain augmentation coefficients alpha and beta = not reported (random scaling/shifting ranges)
Equation 1 perturbs stain and concentration matrices; the ranges of the random coefficients control augmentation strength and are not specified, so the exact recipe is unreproducible.
Ensemble composition = not reported
Inference ensembles 'the predictions of the best found models' but the paper does not say how many models, how selected, or how predictions are aggregated.

assumptions (4)

domain assumption NuClick, run from point annotations, produces accurate mitosis segmentation masks.
Used as ground truth for training the segmentation network; errors in auto-extracted masks propagate into the learned model. No manual review of masks is described.
domain assumption Vahadane stain separation correctly estimates stain and concentration matrices for MIDOG++ images.
The entire stain augmentation pipeline (Equation 1) rests on this decomposition being reliable across scanners and tissue types.
domain assumption The loss on the validation domain is a legitimate model-selection criterion in a domain-generalization evaluation.
Explicitly used in the protocol; strict domain generalization requires target-domain labels not to be used for any decision, so this assumption is contested and is the weakest premise.
domain assumption Tracing mitosis detection to segmentation, followed by dilation and connected-component detection, preserves the ability to count mitoses.
Post-processing converts masks to points; the dilation step suppresses double detection of daughter cells but no ablation shows its effect on F1.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Mitosis detection in domain shift scenarios: a Mamba-based approach." pith.science (2026). https://pith.science/paper/64MBSU6R

@misc{pith2026250821033,
  author       = {Pith},
  title        = {Pith review of: Mitosis detection in domain shift scenarios: a Mamba-based approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/64MBSU6R}},
  note         = {Machine review of arXiv:2508.21033}
}

read the original abstract

Mitosis detection in histopathology images plays a key role in tumor assessment. Although machine learning algorithms could be exploited for aiding physicians in accurately performing such a task, these algorithms suffer from significative performance drop when evaluated on images coming from domains that are different from the training ones. In this work, we propose a Mamba-based approach for mitosis detection under domain shift, inspired by the promising performance demonstrated by Mamba in medical imaging segmentation tasks. Specifically, our approach exploits a VM-UNet architecture for carrying out the addressed task, as well as stain augmentation operations for further improving model robustness against domain shift. Our approach has been submitted to the track 1 of the MItosis DOmain Generalization (MIDOG) challenge. Preliminary experiments, conducted on the MIDOG++ dataset, show large room for improvement for the proposed method.

Figures

Figures reproduced from arXiv: 2508.21033 by the authors.

**Figure 1.** VM-UNet architecture applying random stain pertubation to the training images: specifically, first, the stain S and the concentration C matrices of each image I0 are estimated, through the V ahadane method (10). Then, scaling and shifting operations with random coefficients α and β are applied to the stain matrix and on the concentration matrix respectively, to obtain new altered matrices. These altered matrices … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 10 canonical work pages

[1]

Bertram, Robert Klopfleisch, Natalie ter Hoeve, Francesco Ciompi, Frauke Wilm, Christian Marzahl, Taryn A

Marc Aubreville, Nikolas Stathonikos, Christof A. Bertram, Robert Klopfleisch, Natalie ter Hoeve, Francesco Ciompi, Frauke Wilm, Christian Marzahl, Taryn A. Donovan, Andreas Maier, Jack Breen, Nishant Ravikumar, Y oujin Chung, Jinah Park, Ramin Nateghi, Fat- taneh Pourakpour, Rutger H.J. Fick, Saima Ben Hadj, Mostafa Jahanifar, Adam Shephard, Jakob Dexl, ...

work page arXiv 2023
[2]

A generalizable and robust deep learning algorithm for mitosis detection in multicenter breast histopathological images.Medical Image Analysis, 84:102703, 2023

Xiyue Wang, Jun Zhang, Sen Y ang, Jingxi Xiang, Feng Luo, Minghui Wang, Jing Zhang, Wei Y ang, Junzhou Huang, and Xiao Han. A generalizable and robust deep learning algorithm for mitosis detection in multicenter breast histopathological images.Medical Image Analysis, 84:102703, 2023

work page 2023
[3]

Mitosis detection, fast and slow: robust and efficient detection of mitotic figures

Mostafa Jahanifar, Adam Shephard, Neda Zamanitajeddin, Simon Graham, Shan E Ahmed Raza, Fayyaz Minhas, and Nasir Rajpoot. Mitosis detection, fast and slow: robust and efficient detection of mitotic figures. Medical Image Analysis, 94:103132, 2024

work page 2024
[4]

Bertram, Katharina Breininger, Dominik Hirling, Peter Horvath, Nikolas Stathonikos, and Mitko Veta

Jonas Ammeling, Marc Aubreville, Sweta Banerjee, Christof A. Bertram, Katharina Breininger, Dominik Hirling, Peter Horvath, Nikolas Stathonikos, and Mitko Veta. Mitosis Do- main Generalization Challenge 2025. Zenodo, March 2025. doi: 10.5281/zenodo.15077361

doi:10.5281/zenodo.15077361 2025
[5]

Vmamba: Visual state space model

Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Y aowei Wang, Qixiang Y e, Jianbin Jiao, and Yunfan Liu. Vmamba: Visual state space model. Advances in neural information processing systems, 37:103031–103063, 2024

work page 2024
[6]

Mamba-sea: A mamba-based framework with global-to-local sequence augmentation for generalizable medical image segmentation

Zihan Cheng, Jintao Guo, Jian Zhang, Lei Qi, Luping Zhou, Yinghuan Shi, and Y ang Gao. Mamba-sea: A mamba-based framework with global-to-local sequence augmentation for generalizable medical image segmentation. IEEE Transactions on Medical Imaging, 2025

work page 2025
[7]

Nuclick: a deep learning framework for interactive segmentation of microscopic images

Navid Alemi Koohbanani, Mostafa Jahanifar, Neda Zamani Tajadin, and Nasir Rajpoot. Nuclick: a deep learning framework for interactive segmentation of microscopic images. Medical Image Analysis, 65:101771, 2020

work page 2020
[8]

Vm-unet: Vision mamba unet for medi- cal image segmentation

Jiacheng Ruan, Jincheng Li, and Suncheng Xiang. Vm-unet: Vision mamba unet for medi- cal image segmentation. arXiv preprint arXiv:2402.02491, 2024

arXiv 2024

Show all 14 references

[9]

Imagenet: Constructing a large-scale image database

Li Fei-Fei, Jia Deng, and Kai Li. Imagenet: Constructing a large-scale image database. Journal of vision, 9(8):1037–1037, 2009

2009
[10]

Structure-preserving color normalization and sparse stain separation for histological im- ages

Abhishek Vahadane, Tingying Peng, Amit Sethi, Shadi Albarqouni, Lichao Wang, Maxi- milian Baust, Katja Steiger, Anna Melissa Schlitter, Irene Esposito, and Nassir Navab. Structure-preserving color normalization and sparse stain separation for histological im- ages. IEEE transa...

1962
[11]

V-net: Fully convolutional neural networks for volumetric medical image segmentation

Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi. V-net: Fully convolutional neural networks for volumetric medical image segmentation. In 2016 fourth international confer- ence on 3D vision (3DV), pages 565–571. Ieee, 2016

2016
[12]

Focal loss for dense object detection

Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision , pages 2980–2988, 2017

2017
[13]

Dono- van, Samir Jabari, Mitko Veta, Jonathan Ganz, Jonas Ammeling, Paul J

Marc Aubreville, Frauke Wilm, Nikolas Stathonikos, Katharina Breininger, Taryn A. Dono- van, Samir Jabari, Mitko Veta, Jonathan Ganz, Jonas Ammeling, Paul J. Van Diest, Robert Klopfleisch, and Christof A. Bertram. A comprehensive multi-domain dataset for mitotic figure detecti...

2023 doi
[14]

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 Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015. Approach for MIDOG 2025 bioRχiv | 3

2015

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Bertram, Robert Klopfleisch, Natalie ter Hoeve, Francesco Ciompi, Frauke Wilm, Christian Marzahl, Taryn A

Marc Aubreville, Nikolas Stathonikos, Christof A. Bertram, Robert Klopfleisch, Natalie ter Hoeve, Francesco Ciompi, Frauke Wilm, Christian Marzahl, Taryn A. Donovan, Andreas Maier, Jack Breen, Nishant Ravikumar, Y oujin Chung, Jinah Park, Ramin Nateghi, Fat- taneh Pourakpour, Rutger H.J. Fick, Saima Ben Hadj, Mostafa Jahanifar, Adam Shephard, Jakob Dexl, ...

work page arXiv 2023

[2] [2]

A generalizable and robust deep learning algorithm for mitosis detection in multicenter breast histopathological images.Medical Image Analysis, 84:102703, 2023

Xiyue Wang, Jun Zhang, Sen Y ang, Jingxi Xiang, Feng Luo, Minghui Wang, Jing Zhang, Wei Y ang, Junzhou Huang, and Xiao Han. A generalizable and robust deep learning algorithm for mitosis detection in multicenter breast histopathological images.Medical Image Analysis, 84:102703, 2023

work page 2023

[3] [3]

Mitosis detection, fast and slow: robust and efficient detection of mitotic figures

Mostafa Jahanifar, Adam Shephard, Neda Zamanitajeddin, Simon Graham, Shan E Ahmed Raza, Fayyaz Minhas, and Nasir Rajpoot. Mitosis detection, fast and slow: robust and efficient detection of mitotic figures. Medical Image Analysis, 94:103132, 2024

work page 2024

[4] [4]

Bertram, Katharina Breininger, Dominik Hirling, Peter Horvath, Nikolas Stathonikos, and Mitko Veta

Jonas Ammeling, Marc Aubreville, Sweta Banerjee, Christof A. Bertram, Katharina Breininger, Dominik Hirling, Peter Horvath, Nikolas Stathonikos, and Mitko Veta. Mitosis Do- main Generalization Challenge 2025. Zenodo, March 2025. doi: 10.5281/zenodo.15077361

doi:10.5281/zenodo.15077361 2025

[5] [5]

Vmamba: Visual state space model

Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Y aowei Wang, Qixiang Y e, Jianbin Jiao, and Yunfan Liu. Vmamba: Visual state space model. Advances in neural information processing systems, 37:103031–103063, 2024

work page 2024

[6] [6]

Mamba-sea: A mamba-based framework with global-to-local sequence augmentation for generalizable medical image segmentation

Zihan Cheng, Jintao Guo, Jian Zhang, Lei Qi, Luping Zhou, Yinghuan Shi, and Y ang Gao. Mamba-sea: A mamba-based framework with global-to-local sequence augmentation for generalizable medical image segmentation. IEEE Transactions on Medical Imaging, 2025

work page 2025

[7] [7]

Nuclick: a deep learning framework for interactive segmentation of microscopic images

Navid Alemi Koohbanani, Mostafa Jahanifar, Neda Zamani Tajadin, and Nasir Rajpoot. Nuclick: a deep learning framework for interactive segmentation of microscopic images. Medical Image Analysis, 65:101771, 2020

work page 2020

[8] [8]

Vm-unet: Vision mamba unet for medi- cal image segmentation

Jiacheng Ruan, Jincheng Li, and Suncheng Xiang. Vm-unet: Vision mamba unet for medi- cal image segmentation. arXiv preprint arXiv:2402.02491, 2024

arXiv 2024