Pith. sign in

REVIEW 5 major objections 5 minor 31 references

Solutions for Mitotic Figure Detection and Atypical Classification in MIDOG 2025

T0 review · 5 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 Two-stage cascade detects mitotic figures at F1 0.8432

desk verdict A plausible challenge report undone by an inconsistent baseline in its main detection claim; the atypical classification section is more solid. read the letter →

arxiv 2509.02597 v1 pith:Q5H6GSYH submitted 2025-08-29 eess.IV cs.CV

Shuting Xu , Runtong Liu , Zhixuan Chen , Junlin Hou , Hao Chen This is my paper
classification eess.IVcs.CV
keywords mitoticfiguredetectionatypicalmitosisclassificationcomputationalpathologytwo-stagemodelensembleFCOSConvNeXtwhole-slidehistology
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

The paper argues that mitotic figure detection in cancer pathology improves when a fast, sensitive detector proposes candidates and a separate image classifier then filters them, removing lookalike cells that confuse single-stage models. On a validation split of the challenge's multi-domain dataset, adding a ResNet50-and-EfficientNet ensemble classifier raised F1 from 0.803 to 0.8432, at the cost of lower average precision. For the companion task of distinguishing atypical from normal mitotic figures, the paper shows that averaging predictions across several ConvNeXt-family models reaches 0.8794 balanced accuracy in cross-validation. The work is a challenge solution: its claims are measured on defined training and validation splits, not a new biological mechanism.

What carries the argument

The two-stage detection-classification cascade: stage one is the FCOS anchor-free detector predicting centers, box offsets, and objectness; stage two crops each positive candidate to a patch and runs a ResNet50/EfficientNet ensemble classifier that keeps only confirmed candidates. The companion machinery is probability averaging across independently trained models for atypical/normal classification, with the ConvNeXt+CBAM family as the strongest single backbones.

What would settle it

Run the same detection pipeline under repeated cross-validation on the MIDOG++ dataset, reporting F1 and AP as distributions, and plot the precision-recall curves of the detector alone versus detector-plus-classifier at matched recall. If the cascade's F1 advantage does not persist across folds or shifts with threshold, the claimed benefit is not robust.

Watch

Extended reading notes

Core claim

The central claim is that detection and classification should be separated: an anchor-free FCOS detector is tuned for high recall, and a dedicated convolutional classifier re-examines each proposed patch to suppress impostor cells such as apoptotic bodies and lymphocytes. The paper reports that this cascade raises F1 from 0.803 to 0.8432 on its validation split while noting AP decreases from 0.842 to 0.787, and that ensembling two classifiers outperforms either alone. For atypical classification, the paper finds convolutional models, especially the ConvNeXt family with attention, outperform a Vision Transformer on limited data, and that ensembling the top convolutional models gives the best

Load-bearing premise

The detection results rely on one fixed training/validation split with no error bars, so the F1 boost from the added classifier could be an artifact of that split or the score threshold rather than a stable gain.

Editorial extensions

If this is right

  • If the cascade generalizes, single-stage histology detectors can be improved by a separate patch-level filter without retraining the detector.
  • The F1 gain indicates false positives from morphologically similar non-mitotic cells are the bottleneck, and a classifier trained on curated positives and negatives directly attacks that bottleneck.
  • The finding that adding canine breast-cancer data helps but canine mast-cell-tumor data hurts suggests domain composition of training data is a first-order choice.
  • Ensembling convolutional backbones provides more reliable atypical/normal scores than any single network, which matters for clinical grading where false negatives are costly.

Reading between the lines

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

  • Editorially, the F1 improvement may partly reflect threshold selection: AP falls from 0.842 to 0.787 after the cascade, so the F1 gain could shrink or reverse at matched operating points; a precision-recall curve comparison would settle it.
  • Editorially, all detection comparisons rest on one 80/20 split; repeated splits or cross-validation would be needed to confirm that the cascade gain is not split-specific.
  • Editorially, the same cascade idea could be tested on other small-object detection tasks in whole-slide histology, such as detecting other rare cell events, where a recall-first detector plus a discriminative patch classifier is a natural fit.
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.

Referee Report

5 major / 5 minor

Summary. This manuscript reports the authors' solutions to the two MIDOG 2025 challenge tasks. For mitotic figure detection, they propose a two-stage pipeline: an FCOS detector first proposes candidate regions, and a secondary classifier (ResNet-50 or an ensemble of ResNet-50 and EfficientNet-B2) re-evaluates each candidate to reduce false positives. They report that this cascade improves F1 from 0.803 to 0.8432 on a validation split of MIDOG++. For atypical mitosis classification, they compare several CNN and ViT backbones using 4-fold cross-validation and then ensemble the ConvNeXt family, reporting a balanced accuracy of 0.8794 in cross-validation and 0.86 on the official preliminary evaluation. The paper is written as a challenge-participation report, with implementation details and public datasets.

Significance. If the detection result were properly supported, the main contribution would be a practical demonstration that a detection-classification cascade can boost mitotic-figure detection F1 beyond a strong one-stage detector, which is of interest to the computational pathology community. The atypical classification part is more carefully evaluated: it uses 4-fold cross-validation with standard deviations, compares several modern architectures, and reports an official preliminary evaluation score. The paper also contributes useful details about training data combinations (MIDOG++, CMC, CMCCT) and an honest limitation that no external validation was used during development. However, the central detection claim is currently undermined by an internal inconsistency in the baseline comparison and by the absence of operating-point and variability analyses.

major comments (5)
  1. [Table 4 and 'Validation of detection-classification framework'] The baseline row in Table 4 (F1=0.803, AP=0.842) is identical to the 'CMC + MIDOG++' row in Table 2, while the text states the baseline was 'trained on MIDOG++'. Table 2 reports MIDOG++-only as F1=0.794, AP=0.782. If the baseline actually used CMC+MIDOG++, then the F1 gain from the cascade (0.803 to 0.8432) is confounded with the gain from adding CMC training data; if the baseline was indeed MIDOG++-only, the table numbers are inconsistent. This must be corrected and the comparison re-run before the central detection claim can be evaluated.
  2. [Table 4, F1 vs AP] The cascade increases F1 from 0.803 to 0.8432 but decreases AP from 0.842 to 0.787. F1 is computed at a single decision threshold, and the classifier threshold is not reported. A re-ranking cascade can improve F1 at one threshold while degrading the precision-recall ranking, as AP indicates here. Without reporting the threshold, precision-recall curves, or a matched-recall comparison, the F1 improvement alone does not demonstrate 'enhanced detection robustness'. Please provide operating-point details and a threshold-independent analysis (e.g., PR curves or AP with confidence intervals).
  3. [Tables 1-4, single split] All detection results are based on one fixed 80/20 split, with no cross-validation, repeated runs, or error bars. The classification patch split (7:1:2) in Table 3 also appears to be a single split. Given the small F1 differences involved (e.g., 0.794 vs 0.803) and the threshold sensitivity noted above, the manuscript should report variability across at least a few seeds or split folds, or explicitly state that these are single-run values and avoid making strong comparative claims from them.
  4. ['Validation of different classification models' and Table 4] The classifier in the cascade is trained on 50x50 patches extracted from MIDOG++ with a 7:1:2 split, while the detector is validated on an 80/20 split of MIDOG++. The paper does not state whether the classifier training/validation patches were separated from the detector validation split. If the classifier saw patches from images that also appear in the detector's validation set, this would leak information and inflate the reported F1. Please clarify the relationship between the patch split and the detection split, and if necessary re-evaluate with strictly disjoint splits.
  5. ['Preliminary evaluation' and 'Model ensemble'] The final atypical-classification ensemble is described only as 'the family of ConvNeXt models'. It is not specified which models (ConvNeXt_large, ConvNeXt_base, ConvNeXt_base+CBAM?) are included, how many folds contributed, or how the ensemble was selected. Moreover, the cross-validation table used for model selection is also presented as the main result, which risks selection bias. The official preliminary evaluation score (0.86) is reported without confidence intervals or details about the evaluation set size. Please specify the ensemble construction and report the official evaluation with as much metadata as available.
minor comments (5)
  1. [Introduction] The sentence ending 'similarity to impostor cells (e.g., apoptotic cells or lymphocytes).' is grammatically incomplete; a verb or clause is missing.
  2. [Table 3] Table 3 reports no standard deviations or confidence intervals, whereas Table 5 does. Adding variability estimates to Table 3 would help interpretation.
  3. [Table 5] EfficientNetB1's balanced accuracy is reported as 0.8325 ± 0.0000, which suggests identical splits or a rounding artifact; please verify.
  4. [References] Several references contain formatting artifacts, e.g., 'Y equan Bie' instead of 'Yequan Bie'.
  5. [Implementation details] The text says 'one NVIDIA GeForce RTX 4090 GPUs' but singular/plural agreement is inconsistent.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: reported results are empirical validations, not derivations from their own inputs.

full rationale

The paper's detection and atypical-classification sections report empirical comparisons on fixed validation splits; no equation defines a target quantity in terms of the result being predicted. The two-stage detector-classifier F1 is obtained by running a detector and a classifier, not by fitting the F1 itself. The single-split design and the Table 4 baseline discrepancy (text says 'trained on MIDOG++' but the quoted F1/AP match the CMC+MIDOG++ row of Table 2) are statistical/reporting concerns, not circularity. Self-citations are for standard techniques (e.g., [28] for ensemble averaging) and are not load-bearing. No uniqueness theorem, ansatz, or renaming is invoked. Therefore no circular step can be exhibited with the required quote-and-reduction standard.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims rest on three hand-chosen design parameters (threshold, split, ensemble choice), plus the assumption that the challenge labels and metrics are as described. No new physical or biological entities are introduced.

free parameters (3)
  • Classifier decision threshold in the two-stage cascade = not reported
    The final detection output depends on the score threshold at which the secondary classifier rejects FCOS candidates. This threshold directly controls the F1/AP trade-off observed in Table 4, and is fitted to the validation split but never disclosed.
  • Single 80/20 split for detection and 7:1:2 split for classifier patches = 8:2, 7:1:2
    Results in Tables 1-4 come from one random split. No cross-validation or confidence intervals are given, so the reported F1/AP values depend on the specific split.
  • Ensemble composition for atypical classification = ConvNeXt family, equal-weight soft voting
    The final ensemble was chosen after comparing Table 5 cross-validation results, so the reported balanced accuracy (0.8794) is on the same data used to select the ensemble.
assumptions (4)
  • domain assumption MIDOG++/MIDOG 2025 annotations and auxiliary dataset labels are correct
    All experiments use challenge ground truth (refs 12, 19, 20). If labels are noisy, the reported metrics are affected. Invoked in Dataset sections.
  • domain assumption The evaluation metrics match the challenge definitions
    The paper uses F1, AP, and balanced accuracy but does not define the AP IoU threshold or whether the official challenge rank uses F1 or AP. Table 1 and Table 2 give different AP values for the same FCOS model, suggesting inconsistent metric definitions.
  • domain assumption No data leakage between detection and classification training patches
    The paper does not state whether the 50x50 patches used to train the classifier are extracted from the same whole-slide images and the same split used to train FCOS. Overlap could inflate cascade performance. Assumed implicitly in the two-stage evaluation.
  • domain assumption Standard deep learning training assumptions
    Optimization, augmentation, and generalization behavior of PyTorch training with the given hyperparameters is taken as standard. No formal guarantees are claimed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Solutions for Mitotic Figure Detection and Atypical Classification in MIDOG 2025." pith.science (2026). https://pith.science/paper/Q5H6GSYH

@misc{pith2026250902597,
  author       = {Pith},
  title        = {Pith review of: Solutions for Mitotic Figure Detection and Atypical Classification in MIDOG 2025},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q5H6GSYH}},
  note         = {Machine review of arXiv:2509.02597}
}
read the original abstract

Deep learning has driven significant advances in mitotic figure analysis within computational pathology. In this paper, we present our approach to the Mitosis Domain Generalization (MIDOG) 2025 Challenge, which consists of two distinct tasks, i.e., mitotic figure detection and atypical mitosis classification. For the mitotic figure detection task, we propose a two-stage detection-classification framework that first localizes candidate mitotic figures and subsequently refines the predictions using a dedicated classification module. For the atypical mitosis classification task, we employ an ensemble strategy that integrates predictions from multiple state-of-the-art deep learning architectures to improve robustness and accuracy. Extensive experiments demonstrate the effectiveness of our proposed methods across both tasks.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 15 canonical work pages

  1. [1]

    Periphery-aware covid-19 diagnosis with contrastive representation enhancement

    Junlin Hou, Jilan Xu, Longquan Jiang, Shanshan Du, Rui Feng, Yuejie Zhang, Fei Shan, and Xiangyang Xue. Periphery-aware covid-19 diagnosis with contrastive representation enhancement. Pattern Recognition, 118:108005, 2021

    work page 2021

  2. [2]

    Cross-field transformer for diabetic retinopathy grading on two-field fundus images

    Junlin Hou, Jilan Xu, Fan Xiao, Rui-Wei Zhao, Yuejie Zhang, Haidong Zou, Lina Lu, Wenwen Xue, and Rui Feng. Cross-field transformer for diabetic retinopathy grading on two-field fundus images. In 2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 985–990. IEEE, 2022

    work page 2022

  3. [3]

    Diabetic retinopathy grading with weakly-supervised lesion priors

    Junlin Hou, Fan Xiao, Jilan Xu, Rui Feng, Yuejie Zhang, Haidong Zou, Lina Lu, and Wen- wen Xue. Diabetic retinopathy grading with weakly-supervised lesion priors. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE, 2023

    work page 2023

  4. [4]

    Self-explainable ai for medical image analysis: A survey and new outlooks

    Junlin Hou, Sicen Liu, Y equan Bie, Hongmei Wang, Andong Tan, Luyang Luo, and Hao Chen. Self-explainable ai for medical image analysis: A survey and new outlooks. arXiv preprint arXiv:2410.02331, 2024

    arXiv 2024

  5. [5]

    Concept-attention whitening for interpretable skin lesion diagnosis

    Junlin Hou, Jilan Xu, and Hao Chen. Concept-attention whitening for interpretable skin lesion diagnosis. In International Conference on Medical Image Computing and Computer- Assisted Intervention, pages 113–123. Springer, 2024

    work page 2024

  6. [6]

    Qmix: Quality-aware learning with mixed noise for robust retinal disease diagnosis

    Junlin Hou, Jilan Xu, Rui Feng, and Hao Chen. Qmix: Quality-aware learning with mixed noise for robust retinal disease diagnosis. IEEE Transactions on Medical Imaging, 2025

    work page 2025

  7. [7]

    Data-efficient histopathology image analysis with deformation representation learning

    Jilan Xu, Junlin Hou, Yuejie Zhang, Rui Feng, Chunyang Ruan, Tao Zhang, and Weiguo Fan. Data-efficient histopathology image analysis with deformation representation learning. In 2020 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) , pages 857–864. IEEE, 2020

    work page 2020

  8. [8]

    Segment anything in pathology images with natural language

    Zhixuan Chen, Junlin Hou, Liqi Lin, Yihui Wang, Y equan Bie, Xi Wang, Y anning Zhou, Ronald Cheong Kin Chan, and Hao Chen. Segment anything in pathology images with natural language. arXiv preprint arXiv:2506.20988, 2025

    arXiv 2025

Show all 31 references
  1. [9]

    Mitosis detection in breast cancer histology images with deep neural networks

    Dan C Cire¸ san, Alessandro Giusti, Luca M Gambardella, and Jürgen Schmidhuber. Mitosis detection in breast cancer histology images with deep neural networks. In International conference on medical image computing and computer-assisted intervention , pages 411–

  2. [10]

    Mitosis detection in breast cancer histology images via deep cascaded networks

    Hao Chen, Qi Dou, Xi Wang, Jing Qin, and Pheng Heng. Mitosis detection in breast cancer histology images via deep cascaded networks. In Proceedings of the AAAI conference on artificial intelligence, volume 30, 2016

    2016

  3. [11]

    Mitosis detection techniques in h&e stained breast cancer pathological images: A comprehensive review

    Xipeng Pan, Yinghua Lu, Rushi Lan, Zhenbing Liu, Zujun Qin, Huadeng Wang, and Zaiyi Liu. Mitosis detection techniques in h&e stained breast cancer pathological images: A comprehensive review. Computers & Electrical Engineering, 91:107038, 2021

    2021

  4. [12]

    A comprehensive multi-domain dataset for mitotic figure detection

    Marc Aubreville, Frauke Wilm, Nikolas Stathonikos, Katharina Breininger, Taryn A Donovan, Samir Jabari, Mitko Veta, Jonathan Ganz, Jonas Ammeling, Paul J van Diest, et al. A comprehensive multi-domain dataset for mitotic figure detection. Scientific data, 10(1):484, 2023

    2023

  5. [13]

    A large-scale dataset for mitotic figure assessment on whole slide images of canine cutaneous mast cell tumor

    Christof A Bertram, Marc Aubreville, Christian Marzahl, Andreas Maier, and Robert Klopfleisch. A large-scale dataset for mitotic figure assessment on whole slide images of canine cutaneous mast cell tumor. Scientific data, 6(1):274, 2019

    2019

  6. [14]

    A completely annotated whole slide image dataset of canine breast cancer to aid human breast cancer research

    Marc Aubreville, Christof A Bertram, Taryn A Donovan, Christian Marzahl, Andreas Maier, and Robert Klopfleisch. A completely annotated whole slide image dataset of canine breast cancer to aid human breast cancer research. Scientific data, 7(1):417, 2020

    2020

  7. [15]

    Fcos: Fully convolutional one-stage object detection

    Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9627–9636, 2019

    2019

  8. [16]

    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, pages 770–778, 2016

    2016

  9. [17]

    A method for stochastic optimization

    Diederik Kinga, Jimmy Ba Adam, et al. A method for stochastic optimization. In International conference on learning representations (ICLR), volume 5. California;, 2015

    2015

  10. [18]

    A stochastic approximation method

    Herbert Robbins and Sutton Monro. A stochastic approximation method. The annals of mathematical statistics, pages 400–407, 1951

    1951

  11. [19]

    A dataset of atypical vs normal mitoses classification for midog - 2025, April 2025

    Viktoria Weiss, Sweta Banerjee, Taryn Donovan, Thomas Conrad, Robert Klopfleisch, Jonas Ammeling, Christopher Kaltenecker, Dominik Hirling, Mitko Veta, Nikolas Stathonikos, Peter Horvath, Katharina Breininger, Marc Aubreville, and Christof Bertram. A dataset of atypical vs nor...

    2025

  12. [20]

    His- tologic dataset of normal and atypical mitotic figures on human breast cancer (ami-br)

    Christof A Bertram, Viktoria Weiss, Taryn A Donovan, Sweta Banerjee, Thomas Conrad, Jonas Ammeling, Robert Klopfleisch, Christopher Kaltenecker, and Marc Aubreville. His- tologic dataset of normal and atypical mitotic figures on human breast cancer (ami-br). In BVM Workshop, p...

    2025

  13. [21]

    Mitosis domain generalization in histopathology images–the midog challenge

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

    2022 arXiv

  14. [22]

    Pre- dicting breast tumor proliferation from whole-slide images: the tupac16 challenge

    Mitko Veta, Yujing J Heng, Nikolas Stathonikos, Babak Ehteshami Bejnordi, Francisco Beca, Thomas Wollmann, Karl Rohr, Manan A Shah, Dayong Wang, Mikael Rousson, et al. Pre- dicting breast tumor proliferation from whole-slide images: the tupac16 challenge. Medical image analysi...

    2019

  15. [23]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, G Heigold, S Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on...

    2020

  16. [24]

    Densely connected convolutional networks

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700–4708, 2017

    2017

  17. [25]

    A convnet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Sain- ing Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976–11986, 2022

    2022

  18. [26]

    Cbam: Convolutional block attention module

    Sanghyun Woo, Jongchan Park, Joon-Y oung Lee, and In So Kweon. Cbam: Convolutional block attention module. In Proceedings of the European conference on computer vision (ECCV), pages 3–19, 2018

    2018

  19. [27]

    Efficientnet: Rethinking model scaling for convolutional neural networks

    Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105–6114. PMLR, 2019

    2019

  20. [28]

    Cmc-cov19d: Contrastive mixup classification for covid-19 diagnosis

    Junlin Hou, Jilan Xu, Rui Feng, Yuejie Zhang, Fei Shan, and Weiya Shi. Cmc-cov19d: Contrastive mixup classification for covid-19 diagnosis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 454–461, 2021

    2021

  21. [29]

    Faster r-cnn: Towards real-time object detection with region proposal networks

    Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015

    2015

  22. [30]

    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

  23. [31]

    Aggregated resid- ual transformations for deep neural networks

    Saining Xie, Ross Girshick, Piotr Dollár, Zhuowen Tu, and Kaiming He. Aggregated resid- ual transformations for deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1492–1500, 2017. 4 | Shuting Xu et al. Solutions for MIDOG 2025

    2017

Pith tools

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