A bag of tricks for real-time Mitotic Figure detection

Brian Napora; Christian Marzahl

REVIEW 5 major objections 5 minor 15 references

A bag of tricks for real-time Mitotic Figure detection

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 A compact detector reaches F1 0.81 for mitotic figures across unseen domains

desk verdict Honest engineering note with one real but underspecified external result: F1 0.81 on MIDOG 2025 preliminary, yet the supporting ablations and protocol details are missing. read the letter →

arxiv 2508.19804 v1 pith:UZ5KVV7M submitted 2025-08-27 cs.CV cs.AI

Christian Marzahl , Brian Napora This is my paper

classification cs.CVcs.AI

keywords mitoticfiguredetectiondomaingeneralizationRTMDethistopathologyhardnegativeminingbalancedsamplingsingle-stageobjectMIDOG

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 tries to show that a compact, one-stage object detector can detect mitotic figures across scanners, staining protocols, tissue types, and even species, without sacrificing speed. Using balanced hierarchical sampling over six public datasets, targeted hard-negative mining on necrotic and debris tissue, and an ensemble of EMA-smoothed models, the authors report F1 between 0.78 and 0.84 in grouped 5-fold cross-validation and 0.81 on the preliminary MIDOG 2025 test set. If correct, this means real-time mitotic-figure detection, about 20–30 ms per large image patch, is achievable with a small model that generalizes to unseen domains. The practical payoff is a fast, deployable aid for tumor grading that does not need a large two-stage pipeline.

What carries the argument

The load-bearing mechanism is the three-level hierarchical sampling weight w(x_i)=w_D(d_i)·w_I(m_i)·w_C(c_i). Each factor is an inverse-frequency weight: dataset level, slide level, and class level (MF vs NMF), so every mini-batch draws roughly equally from all six sources, from different slides, and at a 1:1 mitotic/non-mitotic ratio. Around this, RTMDet-S is the compact real-time single-stage detector backbone; hard-negative mining adds extracted patches from necrotic and debris tissue as false-positive examples, and four EMA-smoothed weight copies are ensembled at inference with test-time augmentation.

What would settle it

Ask for the exact slide-level fold assignment and MIDOG submission identifiers, then check whether any specimen appears in both training and test folds; alternatively, rerun the same pipeline on the released final MIDOG 2025 test labels and compare the reproduced F1 to 0.81.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that RTMDet-S, a small single-stage detector with a 64×64 anchor tuned to mitotic figures, reaches F1 0.81 on the MIDOG 2025 preliminary test set and F1 0.78–0.84 in grouped 5-fold cross-validation, outperforming larger RTMDet variants, Swin-L and ConvNeXt backbones, and a two-stage verification pipeline. The authors attribute the result to a three-level sampling weight that balances datasets, slides, and mitotic versus non-mitotic patches, plus HSV augmentation and hard negatives mined from necrotic and debris-labeled regions. They interpret the performance as evidence that careful training-data curation matters more than model capacity for dom

Load-bearing premise

The evaluation assumes the grouped fold split and the MIDOG test set never share a slide or patient with the training data; if they do, the reported F1 range overstates how well the model generalizes.

Editorial extensions

If this is right

At 20–30 ms per 1920×1280 patch on an RTX 3090, a whole slide can be processed in minutes, making mitotic counting practical in a clinical workflow.
Larger models and two-stage verification add cost without improving F1, so a well-trained small one-stage detector is enough for this task.
Hard negatives from necrosis and debris reduce false positives, directly addressing the confounders that plague mitosis counting in real slides.
The negative results for mosaic augmentation, stain normalization, and a second-stage classifier suggest that domain coverage and balanced sampling are the main levers for domain generalization here.

Reading between the lines

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

The same recipe likely extends to other small-object histopathology detection tasks, such as counting other cell types, whenever domain shift and class imbalance dominate; that is an extrapolation, not a paper claim.
The reported F1 of 0.81 rests on the preliminary MIDOG 2025 leaderboard; if the final test set reassigns labels or adds harder domains, the number could shift.
Publishing the exact slide-level fold IDs would let others verify that no specimen overlaps train and test, and would make the 0.78–0.84 range reproducible.
Hard-negative mining could be iterated: run the trained model on unlabeled clinical slides, harvest confident false positives, and retrain, potentially removing the need for manually annotated negative tissue.

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

Take it as an honest engineering note with one real external number. An RTMDet-S ensemble reaches F1 0.81 on the MIDOG 2025 preliminary test set at 20–30 ms per patch, which exceeds the 0.76 F1 of the MIDOG 2022 winner (on an earlier challenge, so the comparison is not apples-to-apples). If the number holds, it is a practical result for clinical MF detection. But the claim that the specific bag of tricks is responsible is mostly asserted rather than shown.

What is genuinely good: the negative findings — mosaic augmentation, stain normalization, and a two-stage detector did not help; larger RTMDet variants did not outdo the small model — are reported fairly and are useful to people working in this area. The method is described concretely enough for a competent group to re-implement the recipe. The hard negative mining is not as circular as the reader worried: the negatives are chosen by external necrosis/debris region labels, so the model's own predictions only propose candidates, not labels. That concern is minor.

Soft spots, in proportion. The three-level balanced sampling is the centerpiece of the 'tricks', yet the only support is 'Our ablations determined this strategy was important' — no numbers, no table. The MIDOG 2025 result is 'preliminary' and comes without protocol details: no metric matching threshold, no per-patch vs per-slide definition, no entry cohort, no definition of 'larger models'. 'Outperforming larger models' is therefore unverifiable from the paper. The grouped 5-fold CV is slide-partitioned by description, which is the right idea, but without specimen IDs or per-fold detail it supports less than the authors claim. No code, weights, or inference script are provided, so the headline F1 cannot be reproduced from the manuscript.

Bottom line: a plausible engineering recipe with one genuinely external but underspecified benchmark result. It deserves peer review — it is short, clear, and the external number makes it worth referee time — but a referee should ask for ablation numbers, challenge protocol details, and a release of code or weights. I would not cite the 0.81 as a settled number until the leaderboard is final or the code appears; the negative findings are worth discussing in a reading group, though.

Referee Report

5 major / 5 minor

Summary. The paper proposes a 'bag of tricks' for real-time mitotic figure detection using the RTMDet-S single-stage detector. The main ingredients are: training on several public mitotic figure datasets spanning scanners, species, and tumor types; a three-level hierarchical sampling strategy balancing datasets, slides, and MF/NMF classes; targeted hard-negative mining from necrotic and debris regions; HSV jitter and geometric augmentation; an ensemble of EMA-smoothed models; and test-time augmentation. The authors report per-fold F1 scores between 0.78 and 0.84 in a grouped 5-fold cross-validation and an F1 of 0.81 on the preliminary MIDOG 2025 test set, with inference at roughly 20-30 ms per 1920x1280 patch on an RTX 3090. They also report negative results for mosaic augmentation, stain normalization, and a two-stage detector.

Significance. If fully substantiated, the result is practically significant: it would show that a compact one-stage detector, trained with carefully balanced multi-domain data and hard negatives, can match or exceed larger models on mitotic figure detection under domain shift while remaining real-time. The negative findings are also potentially useful to the community. However, the paper currently provides almost no quantitative ablation support for its own 'bag of tricks', and the headline external leaderboard result is described with insufficient protocol detail to be verified. The work is therefore best viewed as a challenge-report-level claim that needs substantial additional evidence before it can be accepted as a scientific paper.

major comments (5)

[Abstract and Section IV] The central claim 'F1 of 0.81 on the MIDOG 2025 preliminary test set, outperforming larger models' is not verifiable from the manuscript. The paper does not define the challenge protocol: matching threshold, per-patch vs per-slide F1 computation, number of submitted entries, date of the leaderboard snapshot, or how 'larger models' are defined. No code, weights, predictions, or leaderboard excerpt are provided. Because the leaderboard is 'preliminary', the number is time-indexed and may change. Please provide the full protocol, the exact leaderboard table with ranks and total entries, and release code/weights or a reproducible inference script.
[Section III, Hierarchical Sampling and Section IV, Results] The paper asserts that ablations showed the balanced sampling strategy was 'important', that larger RTMDet variants and Swin/ConvNeXt backbones were 'on par or slightly worse', and that mosaic augmentation, stain normalization, and a second-stage classifier were ineffective. None of these claims is accompanied by a single number, table, or error bar. For a paper whose contribution is a 'bag of tricks', the contribution of each trick is the central scientific content. Please provide an ablation table with mean and per-fold F1, precision, and recall for each component (e.g., sampling strategy, hard negatives, augmentation, EMA ensemble, TTA) and a comparison table of architectures.
[Section IV, Evaluation and Results] The 'grouped, stratified 5-fold cross-validation' is under-specified. 'Grouped' is not defined; the paper does not state the grouping unit (slide, specimen, patient), how stratification was performed, or the number of specimens/slides in each fold. Without assurance that no specimen or patient appears in both training and test folds, the F1 range 0.78-0.84 cannot be interpreted as generalization to unseen slides. Please specify the grouping unit, report the fold compositions, and list per-fold F1, precision, and recall.
[Section III, Hard Negative Mining] The hard-negative mining procedure is self-bootstrapping: the model's own low-confidence detections on necrotic/debris regions are selected as new NMF training examples. While the labels are derived from region annotations, the selection criterion is model-dependent and could reinforce the model's existing false-positive modes. The paper reports no ablation comparing mined vs random negatives and no details on the number of mined patches or potential overlap with the training set. Please provide this information and discuss how the self-selection affects the error profile.
[Section III, Methods] The paper describes RTMDet as a 'one-stage anchor-based detector' and states that an 'appropriate anchor size (64×64 pixels)' was used. In its original formulation, RTMDet is an anchor-free detector. This is a substantive mischaracterization that affects reproducibility. Please clarify whether a non-standard anchor-based head was implemented, or whether 'anchor size' refers to a label-assignment prior; if the standard RTMDet was used, correct the description.

minor comments (5)

[Section III, Eqs. (2)-(3)] The formulas for w_D and w_I include a summation over all datasets/images that does not depend on the current d or m; they reduce to a constant divided by |d| or |m|. If the intent is inverse-frequency weighting, write the normalized form and explain how the product of weights in Eq. (1) is converted into sampling probabilities.
[Section IV, Inference Speed] The inference speed of 20-30 ms per patch is reported without measurement conditions. Please specify batch size, inference precision (FP32/FP16), framework, and whether preprocessing/TTA is included in the timing.
[Datasets and References] The MIDOG 2025 challenge is not cited or described. Please add a citation or URL to the challenge specification, including the definition of Track 1 and the evaluation metric.
[Throughout] Text formatting issues: '112.923 patches' uses a decimal-style separator; the abstract uses 'MItosis' with irregular capitalization; '1920×1280' appears inside a sentence without units. These should be corrected.
[Section III, Augmentation] The augmentation ranges are given as 'e.g. ±10% hue variation, ±20% saturation, ±20% brightness'; 'e.g.' indicates the exact final values are not specified. Please state the exact ranges and any clipping applied.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central F1=0.81 result is an external challenge score, and the proposed training tricks do not reduce to their inputs.

full rationale

The paper's central claim—F1=0.81 on the MIDOG 2025 preliminary test set—is an external benchmark result, not a quantity derived from the method's own equations or fitted parameters. The grouped 5-fold cross-validation (F1 0.78–0.84) is also evaluated on held-out slides, and the paper does not use those folds as training data for the final model. The training techniques described (Eqs. 1–4 balanced sampling, HSV/geometric augmentation, EMA ensembling) are standard and do not encode the evaluation metric or the test set. The hard negative mining step does use a trained RTMDet-S model to propose candidate patches from SPIDER and NCT-CRC-HE-100K, but the resulting examples are labeled as NMF only because they fall in regions annotated as necrotic or debris by those external datasets; the model supplies locations, not ground-truth labels, and these patches are not part of the MIDOG 2025 test set. This is a self-bootstrapping training loop, but it is not a circular prediction chain. Self-citations (refs 1, 2, 4, 6, 7, 8, 9, 12, 13) are dataset sources, prior baselines, or a sampling inspiration; none is a load-bearing uniqueness theorem or a forced ansatz. The paper does omit procedural details for the MIDOG 2025 leaderboard (metric definition, cohort composition, error bars) and does not define what 'grouped' means in the CV, but these are reproducibility/verification gaps, not circular reductions. No equation or fitted parameter is reused as a prediction, so I find no significant circularity.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central claim depends on a set of hand-chosen hyperparameters and data-assumption premises. No new theoretical entities are introduced. The most consequential assumptions are the domain-invariance of MF appearance and the integrity of the data splits.

free parameters (7)

anchor size = 64x64 pixels
Chosen by hand in Section III to match typical MF size at 40x; if MFs vary across scanners, recall drops.
training patch size = 1920x1280 pixels
Chosen in Section III for context around candidates; affects batch composition and model capacity.
training iterations = 25,000
Fixed schedule in Section III; stopping rule is lowest validation loss, but the total is arbitrary.
EMA decay rates = 1e-3, 5e-4, 2.5e-4
Three hand-picked values in Section III; the ensemble of four models depends on these.
hard negative confidence threshold = 0.25
Set in Section III for mining; changing it changes the hard negative pool and thus the trained model.
MF:NMF batch ratio = 1:1
Stated as 'roughly' in Section III; the class-level weight wC encodes this target.
augmentation ranges = ±15 deg rotation; HSV ±10/20/20%; flips 50%
Hand-chosen in Section III; the paper says preliminary experiments showed their benefit but reports no numbers.

assumptions (4)

domain assumption Mitotic figures are roughly constant in size across domains such that a 64x64 anchor at 40x magnification captures them.
Invoked in Section III where the anchor size is set 'to typical MF size at 40x'. If this fails on unseen scanners, the detector cannot localize MFs.
domain assumption Necrotic and debris patches from SPIDER and NCT-CRC-HE-100K contain no true mitotic figures.
Hard negative mining in Section III labels all confident model predictions in those regions as NMF. Any true MFs there would be suppressed as false positives.
domain assumption The 'grouped' cross-validation split preserves slide/patient independence.
Section IV says slides were partitioned to be representative but does not define 'grouped'. If slides from the same specimen appear in both train and test, F1 is inflated.
domain assumption The MIDOG 2025 preliminary test set is an unbiased sample of unseen domains.
The headline F1 of 0.81 rests entirely on this single leaderboard number without the challenge's official scoring description being cited.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A bag of tricks for real-time Mitotic Figure detection." pith.science (2026). https://pith.science/paper/UZ5KVV7M

@misc{pith2026250819804,
  author       = {Pith},
  title        = {Pith review of: A bag of tricks for real-time Mitotic Figure detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UZ5KVV7M}},
  note         = {Machine review of arXiv:2508.19804}
}

read the original abstract

Mitotic figure (MF) detection in histopathology images is challenging due to large variations in slide scanners, staining protocols, tissue types, and the presence of artifacts. This paper presents a collection of training techniques - a bag of tricks - that enable robust, real-time MF detection across diverse domains. We build on the efficient RTMDet single stage object detector to achieve high inference speed suitable for clinical deployment. Our method addresses scanner variability and tumor heterogeneity via extensive multi-domain training data, balanced sampling, and careful augmentation. Additionally, we employ targeted, hard negative mining on necrotic and debris tissue to reduce false positives. In a grouped 5-fold cross-validation across multiple MF datasets, our model achieves an F1 score between 0.78 and 0.84. On the preliminary test set of the MItosis DOmain Generalization (MIDOG) 2025 challenge, our single-stage RTMDet-S based approach reaches an F1 of 0.81, outperforming larger models and demonstrating adaptability to new, unfamiliar domains. The proposed solution offers a practical trade-off between accuracy and speed, making it attractive for real-world clinical adoption.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

15 extracted references · 8 canonical work pages

[1]

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 arXiv:2204.03742, 2022

arXiv 2022
[2]

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),

2022
[3]

Ahmed Raza, Fayyaz Minhas, and Nasir Rajpoot

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. ISSN 1361-

work page 2024
[4]

Domain adver- sarial retinanet as a reference algorithm for the mitosis domain generalization challenge

Frauke Wilm, Christian Marzahl, Katharina Breininger, and Marc Aubreville. Domain adver- sarial retinanet as a reference algorithm for the mitosis domain generalization challenge. In Marc Aubreville, David Zimmerer, and Mattias Heinrich, editors, Biomedical Image Regis- tration, Domain Generalisation and Out-of-Distribution Analysis, pages 5–13, Cham, 202...

work page 2022
[5]

RTMDet: An Empirical Study of Designing Real-Time Object Detec- tors, December 2022

Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Y anyi Liu, Shilong Zhang, and Kai Chen. RTMDet: An Empirical Study of Designing Real-Time Object Detec- tors, December 2022

work page 2022
[6]

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 detection. Scientific Data , 10(1):484, Jul 2023. ISSN 2052-4463. doi: 10.1038/s41597-0...

doi:10.1038/s41597-023-02327-4 2023
[7]

Bertram, Marc Aubreville, Christian Marzahl, Andreas Maier, and Robert Klopfleisch

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, Nov 2019. ISSN 2052-4463. doi: 10.1038/s41597-019-0290-4

doi:10.1038/s41597-019-0290-4 2019
[8]

Bertram, Taryn A

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, Nov 2020. ISSN 2052-4463. doi: 10.1038/s41597-020-00756-z

doi:10.1038/s41597-020-00756-z 2020

Show all 15 references

[9]

Bertram, Mitko Veta, Christian Marzahl, Nikolas Stathonikos, Andreas Maier, Robert Klopfleisch, and Marc Aubreville

Christof A. Bertram, Mitko Veta, Christian Marzahl, Nikolas Stathonikos, Andreas Maier, Robert Klopfleisch, and Marc Aubreville. Are pathologist-defined labels reproducible? com- parison of the tupac16 mitotic figure dataset with an alternative set of labels. In Jaime Cardoso,...

2020
[10]

Spider: A comprehensive multi- organ supervised pathology dataset and baseline models, 2025

Dmitry Nechaev, Alexey Pchelnikov, and Ekaterina Ivanova. Spider: A comprehensive multi- organ supervised pathology dataset and baseline models, 2025

2025
[11]

Predicting survival from colorectal cancer histology slides using deep learning: A retrospec- tive multicenter study

Jakob Nikolas Kather, Johannes Krisam, Pornpimol Charoentong, Tom Luedde, Esther Her- pel, Cleo-Aron Weis, Timo Gaiser, Alexander Marx, Nektarios A Valous, Dyke Ferber, et al. Predicting survival from colorectal cancer histology slides using deep learning: A retrospec- tive mu...

2019
[12]

Bertram, Jennifer Maier, Christian Bergler, Christine Kröger, Jörn Voigt, Katharina Breininger, Robert Klopfleisch, and Andreas Maier

Christian Marzahl, Marc Aubreville, Christof A. Bertram, Jennifer Maier, Christian Bergler, Christine Kröger, Jörn Voigt, Katharina Breininger, Robert Klopfleisch, and Andreas Maier. EXACT: a collaboration toolset for algorithm-aided annotation of images with annotation versio...

2021 doi
[13]

Bertram, Jason Stayt, Anne-Katherine Jasensky, Florian Bartenschlager, Marco Fragoso-Garcia, Ann K

Christian Marzahl, Marc Aubreville, Christof A. Bertram, Jason Stayt, Anne-Katherine Jasensky, Florian Bartenschlager, Marco Fragoso-Garcia, Ann K. Barton, Svenja Elsemann, Samir Jabari, Jens Krauth, Prathmesh Madhu, Jörn Voigt, Jenny Hill, Robert Klopfleisch, and Andreas Maie...

2020 doi
[2022]

doi: 10.5281/zenodo.6362337

doi
[8415]

doi: https://doi.org/10.1016/j.media.2024.103132

2024

Pith tools

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

[1] [1]

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 arXiv:2204.03742, 2022

arXiv 2022

[2] [2]

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),

2022

[3] [3]

Ahmed Raza, Fayyaz Minhas, and Nasir Rajpoot

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. ISSN 1361-

work page 2024

[4] [4]

Domain adver- sarial retinanet as a reference algorithm for the mitosis domain generalization challenge

Frauke Wilm, Christian Marzahl, Katharina Breininger, and Marc Aubreville. Domain adver- sarial retinanet as a reference algorithm for the mitosis domain generalization challenge. In Marc Aubreville, David Zimmerer, and Mattias Heinrich, editors, Biomedical Image Regis- tration, Domain Generalisation and Out-of-Distribution Analysis, pages 5–13, Cham, 202...

work page 2022

[5] [5]

RTMDet: An Empirical Study of Designing Real-Time Object Detec- tors, December 2022

Chengqi Lyu, Wenwei Zhang, Haian Huang, Yue Zhou, Yudong Wang, Y anyi Liu, Shilong Zhang, and Kai Chen. RTMDet: An Empirical Study of Designing Real-Time Object Detec- tors, December 2022

work page 2022

[6] [6]

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 detection. Scientific Data , 10(1):484, Jul 2023. ISSN 2052-4463. doi: 10.1038/s41597-0...

doi:10.1038/s41597-023-02327-4 2023

[7] [7]

Bertram, Marc Aubreville, Christian Marzahl, Andreas Maier, and Robert Klopfleisch

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, Nov 2019. ISSN 2052-4463. doi: 10.1038/s41597-019-0290-4

doi:10.1038/s41597-019-0290-4 2019

[8] [8]

Bertram, Taryn A

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, Nov 2020. ISSN 2052-4463. doi: 10.1038/s41597-020-00756-z

doi:10.1038/s41597-020-00756-z 2020