Challenges and Lessons from MIDOG 2025: A Two-Stage Approach to Domain-Robust Mitotic Figure Detection

Euiseop Song; Jaewoo Park; Jaeyoung Park

REVIEW 4 major objections 5 minor 9 references

Challenges and Lessons from MIDOG 2025: A Two-Stage Approach to Domain-Robust Mitotic Figure Detection

T0 review · 4 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 two-stage mitotic-figure detector finds nearly every mitosis but drowns in look-alike false positives.

desk verdict A credible challenge report with real external scores, but the central lesson about classifier filtering is undermined by an unexplained 20x gap between the internal and official recall numbers. read the letter →

arxiv 2509.02630 v1 pith:EQO55XIW submitted 2025-09-01 eess.IV cs.CV

Euiseop Song , Jaeyoung Park , Jaewoo Park This is my paper

classification eess.IVcs.CV

keywords mitoticfiguredetectiondomaingeneralizationcomputationalpathologyFasterR-CNNensembleclassificationfalsepositivesuppressionMIDOGchallengehistopathology

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 reports the authors' entry in the MIDOG 2025 challenge, which asks algorithms to detect mitotic figures across diverse tissue domains and scanners. Their pipeline pairs a Faster R-CNN candidate detector with an ensemble of three image classifiers meant to remove false positives, and their best submission achieved an F1-score of 0.2237, with recall at 0.9528 but precision at only 0.1267. The core claim is that candidate generation is no longer the bottleneck: the detector essentially finds all true mitoses, but the system is overwhelmed by imposters that look like mitoses, and every attempt at classifier-based filtering made things worse by removing true positives. If correct, this refocuses the field from detection sensitivity to false-positive suppression, and warns that naive sequential filtering can be counterproductive.

What carries the argument

A two-stage pipeline: Faster R-CNN detects candidate bounding boxes, and an ensemble of DenseNet-121, EfficientNet-v2, and InceptionResNet-v2 classifiers averages softmax probabilities to filter candidates. Training combines a pooled dataset (MIDOG++, MITOS-CMC, MITOS-CCMCT) and additional sets (MiDeSeC, ICPR 2012), with D4 rotation, defocus, and RandStainNA stain augmentations. Knowledge distillation from InceptionResNet-v2 improves the DenseNet-121 classifier from F1 0.57 to 0.73, but the ensemble still cannot reconcile candidate-box distribution with classifier training patches.

What would settle it

Evaluate the PooledDB-trained detector separately on each constituent dataset (MIDOG++, MITOS-CMC, MITOS-CCMCT, MiDeSeC, ICPR 2012): if per-source precision varies widely, the pooled precision is a dataset artifact. Separately, feed ground-truth mitotic patches directly into the ensemble classifier: if recall on ground-truth patches is high but recall in the full pipeline is low, the bottleneck is bounding-box misalignment, not classifier discrimination.

Watch

Extended reading notes

Core claim

For the MIDOG 2025 mitotic figure detection task, a Faster R-CNN trained only on the MIDOG++ dataset produced recall of 0.9528 and precision of 0.1267, an F1-score of 0.2237 that was the team's best among six submitted variants. On the internal pooled dataset, detection alone reached recall 0.9842 but precision 0.0578, and adding the three-classifier ensemble collapsed recall to 0.0488 while only modestly improving precision. The authors conclude that high-recall candidate detection is achievable across domains, but that the fundamental limitation is distinguishing true mitotic figures from morphologically similar imposters, and that sequential classifier filtering is not an effective remedy

Load-bearing premise

The detector and classifiers are trained on pooled datasets whose mitotic-figure labeling protocols, scanner profiles, and inclusion criteria are assumed to be equivalent, so if label noise or annotation-style mismatch differs across sources, the low precision could reflect dataset inconsistency rather than a true domain-generalization limit.

Editorial extensions

If this is right

If the central claim holds, future domain-robust mitosis detection systems should prioritize precision-oriented training and false-positive suppression over further recall improvements.
Sequential detection-then-classification pipelines need joint fine-tuning or feature-level integration rather than independent filtering, because the classifier sees different patches than the detector produces.
High recall alone is not a useful clinical metric in this setting; the precision collapse implies that a pathologist would spend most of their time reviewing false positives.
The finding that adding more training data and classifiers decreased performance suggests that naive pooling of heterogeneous mitosis datasets can hurt rather than help domain generalization.
The performance decline of the final submitted variant indicates that overfitting to validation or to specific data sources is a serious risk in challenge-oriented optimization.

Reading between the lines

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

The precision bottleneck might be partly an artifact of cross-dataset labeling differences: if the pooled datasets use different inclusion criteria for 'true' mitoses, the detector is trained on inconsistent labels and the measured precision is understated; per-source evaluation would test this.
The dramatic recall drop after classifier filtering (0.9842 to 0.0488) could reflect bounding-box misalignment: training classifiers on centered annotation patches while testing on detector-generated boxes with imperfect centers may systematically penalize true positives.
A testable extension is to calibrate the ensemble threshold per tissue domain: since precision varied little across submissions but recall varied substantially, a simple per-domain or per-scanner threshold could raise F1 without retraining.
The authors' lesson that 'optimization attempts were counterproductive' may generalize to other small-object detection challenges where the base detector already saturates recall; in such cases, precision-targeted losses or hard-negative mining are more promising than additional classifiers.

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

The one genuinely new thing here is the official MIDOG 2025 test-set scores for this team's submitted variants. Those numbers are externally scored, so they are real measurements, and the paper is refreshingly honest: it says outright that the work was a first challenge attempt, that time constraints caused inconsistent training protocols, and that the final performance was below competitive level. That counts for something. The high-recall/low-precision pattern it reports is also plausible and matches what people already suspect about mitotic detection across domains.

But the main lesson the abstract draws—that two-stage classifier filtering is counterproductive and that false-positive suppression is the binding constraint—does not hold together on the numbers as reported. Section C says the full pipeline drops to recall 0.0488 (F1 0.0646), while the official best submission is reported as recall 0.9528 (F1 0.2237). Those cannot be the same system. If the best submission included the classifier ensemble, the official recall is about twenty times the internal recall. If it did not include the ensemble, then the headline result is not the two-stage pipeline the paper describes. The manuscript never maps the six submitted variants to their scores and never reconciles this gap. That is a load-bearing omission, and it is more serious than the cross-dataset annotation compatibility concern the reader raised. Even if pooling is fine, you still wouldn't know which architecture produced the official result.

There are smaller soft spots: the three classifiers were trained on different datasets with different protocols, and the pooling assumes annotation compatibility across sources without any per-source checks. But those are secondary. The core problem is the attribution gap between the internal and official evaluations.

Who is this for? Future MIDOG contestants and developers of mitotic detection pipelines. It is a data point, not a methodological advance; the two-stage Faster R-CNN plus classifier ensemble is the MDFS-style recipe the paper itself cites. The paper deserves a serious referee rather than a desk reject, because the measurements are real and the flaw is fixable by simply reporting which variant produced which number. But as it stands, I would not cite it for the claim that classifier filtering degrades performance, and I would not treat its lessons as established until the variant-to-score mapping is clarified.

Recommendation: send it to peer review, but the reviewers should push hard on the reconciliation of the internal and official numbers. If that gap remains, the central narrative should be retracted or heavily qualified.

Referee Report

4 major / 5 minor

Summary. The paper reports the authors' participation in the MIDOG 2025 mitotic figure detection challenge. The proposed system is a two-stage pipeline: a Faster R-CNN detector generates candidate boxes, and an ensemble of three CNN classifiers (DenseNet-121, EfficientNet-v2, InceptionResNet-v2) filters false positives. The authors present internal evaluation results on a pooled dataset and official MIDOG 2025 scores for their submissions. Their best official submission achieved F1=0.2237 (recall 0.9528, precision 0.1267) and is attributed to a Faster R-CNN trained solely on MIDOG++. The final submission achieved F1=0.2031. Internal experiments show that the full pipeline drastically reduces recall (0.0488) compared with detection alone (0.9842), leading the authors to conclude that classifier-based false-positive filtering is counterproductive and that false-positive suppression is the main bottleneck in cross-domain mitosis detection.

Significance. If the attribution of the official scores to a specific architecture were clear, this would be a valuable challenge report: it documents a reproducible observation that a high-recall detector can be obtained across domains but at the cost of very low precision, and that sequential classifier filtering can hurt. The paper is transparent about its failures and provides many implementation details and hyperparameters. However, the central lesson is currently not defensible from the reported numbers because the manuscript does not identify which of the six submitted variants produced the official scores, and the internal pipeline metrics are inconsistent with the official best score. The paper's main contribution is a candid but poorly specified negative result that needs substantial revision.

major comments (4)

[Results; Methods C] The best official result (F1=0.2237, R=0.9528, P=0.1267) is attributed to "a Faster R-CNN trained solely on MIDOG++", but it is not stated whether this submission includes the three-classifier ensemble. This is not a pedantic point: Methods C reports that the full pipeline has recall 0.0488 and F1 0.0646, while detection alone has recall 0.9842. Official recall 0.9528 is about 20x the internal full-pipeline recall, so the official best almost certainly did not use the full pipeline as described. If it did, the internal and official numbers are irreconcilable; if it did not, the paper's headline result is not the two-stage method in the abstract. The authors must enumerate the six variants, state which architecture produced each official score, and reconcile the internal and official recall values.
[Methods A; Methods C] The internal evaluation numbers for the detector are inconsistent. Methods A says detection on the PooledDB test set achieved recall 0.9820, precision 0.0578, F1 0.1091. Methods C says "detection alone" achieved F1 0.0831 (recall 0.9842). These two descriptions of the same component cannot both be correct on the same test set. The discrepancy may reflect different test sets (e.g., PooledDB test vs challenge validation), but the manuscript does not define the evaluation split used in Methods C. Without a clear protocol, the claimed degradation from adding the classifier cannot be interpreted.
[Results] The conclusion that "subsequent optimization attempts were counterproductive" is based on comparing the best submission (F1 0.2237) with the final submission (F1 0.2031). No details are given for the other four submissions, nor for what changed in "dataset restructuring and classifier tuning." With only two endpoints and no error bars, the 9.2% difference could be within run-to-run variability or due to specific changes. The paper needs a table listing all six submissions with their configurations and official scores to support this claim.
[Methods A/B] The detector and classifiers are trained on pooled datasets (MIDOG++, MITOS-CMC, MITOS-CCMCT, MiDeSeC, ICPR 2012) under the implicit assumption that mitotic-figure annotations are label-compatible across sources. The paper does not test this assumption (e.g., per-source recall/precision or a domain-adversarial check). Given that the best official submission used a MIDOG++-only detector, the role of pooling in the submitted variants is unclear. This omission leaves open the possibility that the observed low precision is partly an annotation/protocol artifact rather than purely a domain-shift phenomenon.

minor comments (5)

[Abstract] Typo: "omplexity" should be "complexity".
[Footer] The footer "bioRχiv" appears to be an artifact; it should name the actual journal or repository.
[Bibliography] The reference list has typographical issues, e.g., "Y ale" and "Y essica" for author names. Please proofread.
[Methods A/B] No citations or URLs are provided for the datasets MITOS-CMC, MITOS-CCMCT, MiDeSeC, ICPR 2012, and MIDOG++. Please supply references.
[Methods B] It is unclear whether classifier training patches were extracted from the same PooledDB split used for detector training, or whether detection-generated boxes from the validation set were used. Please specify to rule out train/test leakage.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity detected: the paper is an empirical challenge report whose measurements come from an external benchmark, with no fitted-as-prediction step or self-citation chain.

full rationale

The paper reports results from the MIDOG 2025 external challenge. It does not derive a result from an input that presupposes it: detector and classifier performance are measured on held-out challenge data, and the best-submission F1/recall/precision are externally scored. No parameter is fitted to the reported MIDOG 2025 numbers and then 'predicted' as a closely related quantity. The comparison between detection-only and full-pipeline F1 scores is an empirical observation, not an equation in which the conclusion is assumed. The paper contains no self-citations of the authors' prior work, no uniqueness theorem invoked to force a choice, and no renaming of a known result as a new derivation. The internal inconsistency noted between pipeline recall (0.0488) and official recall (0.9528) is a serious correctness/attribution problem, but it is not a circularity: the official numbers are external measurements and the internal numbers are separate evaluations; nothing in the paper defines one in terms of the other. The acknowledged limitations ('Due to time constraints, models were trained on different datasets with varying protocols, which may have introduced methodological inconsistencies') are operational caveats, not circular reasoning. Therefore no circular step can be exhibited, and the circularity score is 0.

Assumptions & free parameters 9 free parameters · 4 assumptions · 0 invented entities

The report introduces no new theoretical objects; its support rests entirely on empirical training choices and benchmark evaluation. Free parameters are hand-set training hyperparameters whose values affect candidate counts and classifier outputs. The main domain assumptions concern cross-dataset label compatibility and transferability of stain augmentation, both load-bearing for the precision numbers that drive the paper's conclusions.

free parameters (9)

patch sampling ratio (foreground:random:imposter) = 5:1:4
Hand-chosen balance for training patch generation in Methods A; directly affects how many imposter candidates the detector learns.
NMS threshold = 0.4
Hand-set in Methods A; determines the candidate count passed to the classifier and hence both recall and precision.
detector training epochs = 150
Fixed in Methods A with no reported selection criterion.
RandStainNA std_hyper = -0.7
Hand-chosen stain augmentation parameter in Methods A; controls the diversity of stain templates and affects domain transfer.
defocus and RandStainNA augmentation probability = p=0.3
Hand-set augmentation odds in Methods A.
knowledge distillation temperature = 4.0
Chosen for DenseNet-121 distillation in Methods B.
KD loss ratio = 0.5
Chosen weight between student cross-entropy and KD loss in Methods B.
EMA decay = 0.9998
Hand-set moving-average coefficient in Methods B.
classifier learning rate = 1e-4
Fixed for the classifiers in Methods B.

assumptions (4)

domain assumption Annotations in the pooled datasets (MIDOG++, MITOS-CMC, MITOS-CCMCT, MiDeSeC, ICPR 2012) are label-compatible and representative of the MIDOG 2025 test distribution.
Invoked in Methods A/B when training detector and classifiers on pooled external data and when interpreting low test precision as a domain-generalization failure.
domain assumption Challenge-provided test labels and F1 evaluation are the ground truth for the central performance claims.
The Results section relies entirely on official MIDOG 2025 scoring; no error bars or alternative metrics are reported.
domain assumption Two-stage detection-then-classification is a valid decomposition of the mitosis detection problem.
Borrowed from the MDFS-style winning pipeline cited in the Introduction; the paper provides no evidence that the decomposition is sound for the adverse-condition track.
domain assumption RandStainNA color-profile statistics based on MIDOG++ transfer to the test domains.
Methods A applies stain augmentation with MIDOG++-derived color profiles; if the profiles mismatch the test stain distribution, precision estimates change.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Challenges and Lessons from MIDOG 2025: A Two-Stage Approach to Domain-Robust Mitotic Figure Detection." pith.science (2026). https://pith.science/paper/EQO55XIW

@misc{pith2026250902630,
  author       = {Pith},
  title        = {Pith review of: Challenges and Lessons from MIDOG 2025: A Two-Stage Approach to Domain-Robust Mitotic Figure Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EQO55XIW}},
  note         = {Machine review of arXiv:2509.02630}
}

read the original abstract

Mitotic figure detection remains a challenging task in computational pathology due to domain variability and morphological complexity. This paper describes our participation in the MIDOG 2025 challenge, focusing on robust mitotic figure detection across diverse tissue domains. We developed a two-stage pipeline combining Faster R-CNN for candidate detection with an ensemble of three classifiers (DenseNet-121, EfficientNet-v2, InceptionResNet-v2) for false positive reduction. Our best submission achieved F1-score 0.2237 (Recall: 0.9528, Precision: 0.1267) using a Faster R-CNN trained solely on MIDOG++ dataset. While our high recall demonstrates effective mitotic figure detection, the critically low precision (12.67%) reveals fundamental challenges in distinguishing true mitoses from morphologically similar imposters across diverse domains. Analysis of six submission variants showed that subsequent optimization attempts were counterproductive, highlighting the omplexity of domain generalization in histopathology. This work provides valuable insights into the practical challenges of developing robust mitotic figure detection algorithms and emphasizes the importance of effective false positive suppression strategies.

Figures

Figures reproduced from arXiv: 2509.02630 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

9 extracted references · 7 canonical work pages

[1]

The 1000 mitoses project: A Consensus-Based international collaborative study on mitotic figures classification

Sherman Lin, Christopher Tran, Ela Bandari, Tommaso Romagnoli, Yueyang Li, Michael Chu, Abinaya S Amirthakatesan, Adam Dallmann, Andrii Kostiukov, Angel Panizo, Anjel- ica Hodgson, Anna R Laury, Antonio Polonia, Ashley E Stueck, Aswathy A Menon, Aurélien Morini, Birsen Özamrak, Caroline Cooper, Celestine Marie G Trinidad, Christian Eisenlöf- fel, Dauda E ...

work page 2024
[2]

Mitosis domain generalization in histopathology images–the midog challenge

Marc Aubreville, Nikolas Stathonikos, Christof A Bertram, Robert Klopleisch, Natalie ter Ho- eve, Francesco Ciompi, Frauke Wilm, Christian Marzahl, Taryn A Donovan, Andreas Maier, et al. Mitosis domain generalization in histopathology images–the midog challenge. arXiv preprint arXiv:2204.03742, 2022

arXiv 2022
[3]

Mitosis domain generalization challenge 2022

Marc Aubreville, Christof Bertram, Katharina Breininger, Samir Jabari, Nikolas Stathonikos, and Mitko Veta. Mitosis domain generalization challenge 2022. In 25th International Con- ference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2022) ,

work page 2022
[4]

RandStainNA: Learning Stain- Agnostic Features from Histology Slides by Bridging Stain Augmentation and Normaliza- tion, page 212–221

Yiqing Shen, Yulin Luo, Dinggang Shen, and Jing Ke. RandStainNA: Learning Stain- Agnostic Features from Histology Slides by Bridging Stain Augmentation and Normaliza- tion, page 212–221. Springer Nature Switzerland, 2022. ISBN 9783031164347. doi: 10.1007/978-3-031-16434-7_21

work page doi:10.1007/978-3-031-16434-7_21 2022
[5]

Weinberger

Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely con- nected convolutional networks, 2018

work page 2018
[6]

Mingxing Tan and Quoc V. Le. Efficientnetv2: Smaller models and faster training, 2021

work page 2021
[7]

Inception-v4, inception- resnet and the impact of residual connections on learning, 2016

Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alex Alemi. Inception-v4, inception- resnet and the impact of residual connections on learning, 2016

work page 2016
[8]

Distilling the knowledge in a neural network, 2015

Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network, 2015. Approach for MIDOG 2025 bioRχiv | 3

work page 2015

Show all 9 references

[2022]

doi: 10.5281/zenodo.6362337

doi

Pith tools

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

[1] [1]

The 1000 mitoses project: A Consensus-Based international collaborative study on mitotic figures classification

Sherman Lin, Christopher Tran, Ela Bandari, Tommaso Romagnoli, Yueyang Li, Michael Chu, Abinaya S Amirthakatesan, Adam Dallmann, Andrii Kostiukov, Angel Panizo, Anjel- ica Hodgson, Anna R Laury, Antonio Polonia, Ashley E Stueck, Aswathy A Menon, Aurélien Morini, Birsen Özamrak, Caroline Cooper, Celestine Marie G Trinidad, Christian Eisenlöf- fel, Dauda E ...

work page 2024

[2] [2]

Mitosis domain generalization in histopathology images–the midog challenge

Marc Aubreville, Nikolas Stathonikos, Christof A Bertram, Robert Klopleisch, Natalie ter Ho- eve, Francesco Ciompi, Frauke Wilm, Christian Marzahl, Taryn A Donovan, Andreas Maier, et al. Mitosis domain generalization in histopathology images–the midog challenge. arXiv preprint arXiv:2204.03742, 2022

arXiv 2022

[3] [3]

Mitosis domain generalization challenge 2022

Marc Aubreville, Christof Bertram, Katharina Breininger, Samir Jabari, Nikolas Stathonikos, and Mitko Veta. Mitosis domain generalization challenge 2022. In 25th International Con- ference on Medical Image Computing and Computer Assisted Intervention (MICCAI 2022) ,

work page 2022

[4] [4]

RandStainNA: Learning Stain- Agnostic Features from Histology Slides by Bridging Stain Augmentation and Normaliza- tion, page 212–221

Yiqing Shen, Yulin Luo, Dinggang Shen, and Jing Ke. RandStainNA: Learning Stain- Agnostic Features from Histology Slides by Bridging Stain Augmentation and Normaliza- tion, page 212–221. Springer Nature Switzerland, 2022. ISBN 9783031164347. doi: 10.1007/978-3-031-16434-7_21

work page doi:10.1007/978-3-031-16434-7_21 2022

[5] [5]

Weinberger

Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely con- nected convolutional networks, 2018

work page 2018

[6] [6]

Mingxing Tan and Quoc V. Le. Efficientnetv2: Smaller models and faster training, 2021

work page 2021

[7] [7]

Inception-v4, inception- resnet and the impact of residual connections on learning, 2016

Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alex Alemi. Inception-v4, inception- resnet and the impact of residual connections on learning, 2016

work page 2016

[8] [8]

Distilling the knowledge in a neural network, 2015

Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network, 2015. Approach for MIDOG 2025 bioRχiv | 3

work page 2015