MIDOG Competition Series

• MIDOG Competition Series is a benchmark platform for mitosis detection in histopathology, emphasizing robust generalization across varied scanners, species, and tumor types.
• The series evolved from addressing scanner-induced variations in breast cancer to including multi-domain challenges and atypical mitotic figure classification.
• Recent editions incorporate diverse ROI selections and advanced methods like ensembling and domain-adversarial training to tackle annotation uncertainty and performance gaps.

The MIDOG Competition Series—MItosis DOmain Generalization—is a sequence of benchmark challenges for mitotic figure analysis in histopathology, organized around the central problem of robust generalization under domain shift. Across its major editions, the series evolved from scanner-focused breast cancer mitosis detection to broader evaluation across tumor types, laboratories, species, and tissue contexts, and in 2025 added a second track for atypical mitotic figure classification. Its clinical motivation is consistent throughout: mitotic activity is a key marker of tumor proliferation, malignancy, grading, and prognosis, yet both manual assessment and algorithmic performance are limited by observer variability and distribution shift (Aubreville et al., 2022, Aubreville et al., 2023, Aubreville et al., 5 Jun 2026).

1. Clinical rationale and conceptual scope

MIDOG was created to study a specific failure mode of computational pathology systems: strong performance on training-like material followed by deterioration when the acquisition or biological domain changes. In the 2021 challenge paper, the decisive component of domain shift was identified as variability caused by using different whole slide scanners; the 2022 overview broadened this to tumor types, laboratories, species, staining/processing conditions, and scanners; the 2025 overview added a further axis, namely contextual diversity within the slide, by evaluating not only hotspot regions but also random and challenging regions rich in hard negatives (Aubreville et al., 2022, Aubreville et al., 2023, Aubreville et al., 5 Jun 2026).

The series is anchored in mitotic figure recognition because mitotic count is clinically important but difficult. The 2021 overview states that recognition of mitotic figures by pathologists is subject to a strong inter-rater bias, while later MIDOG reports repeatedly emphasize that automated methods fail under scanner, stain, laboratory, species, and tumor-type shifts. This makes MIDOG less a generic object-detection benchmark than a domain-generalization benchmark for computational pathology (Aubreville et al., 2022, Shen et al., 28 Aug 2025).

A recurrent theme across the series is that mitosis analysis is not exhausted by localization. MIDOG 2025 formalized this by introducing Track 2, which distinguishes normal mitotic figures (NMFs) from atypical mitotic figures (AMFs). That extension reflects the view, stated in multiple 2025 reports, that atypical mitoses are biologically and prognostically relevant, yet are rarer, more morphologically subtle, and more sensitive to domain shift than ordinary mitotic figure detection (Aubreville et al., 5 Jun 2026, Giedziun et al., 29 Aug 2025).

2. Evolution of the benchmark design

The series expanded in scope through three major stages.

Edition	Benchmark design	Key reported outcome
MIDOG 2021	Training set of 200 cases across four scanning systems; additional 100 cases in test, including two previously unseen scanners	Winning algorithm yielded $F_1 = 0.748$; Top5 ensemble reached 0.773 (Aubreville et al., 2022)
MIDOG 2022	Released training data with 405 tumor cases and 9,501 mitotic figure annotations across six domains; final evaluation on 100 cases across 10 independent tumor domains (Aubreville et al., 2023)	Best performance $F_1 = 0.764$ (Aubreville et al., 2023)
MIDOG 2025	Final test set of 365 ROIs from 122 cases/patients, spanning 12 tumor domains; Track 1 added hotspot, random, and challenging ROIs; Track 2 introduced AMF classification (Aubreville et al., 5 Jun 2026)	Top detection $F_1 = 0.740$; top AMF classification balanced accuracy = 0.908 (Aubreville et al., 5 Jun 2026)

MIDOG 2021 was a MICCAI 2021 challenge centered on breast cancer ROIs under scanner-induced shift. The dataset contained 300 breast cancer cases total, retrospectively curated from the University Medical Center Utrecht archive, with 200 cases for training, 20 for preliminary test, and 80 for final test. The design used six scanners overall, including two scanners present only in testing, and extracted corresponding ROIs across scanners by registration so that the same biological tissue region was evaluated across devices (Aubreville et al., 2022).

MIDOG 2022 deliberately changed the meaning of “domain.” Its training set comprised six domains, including human breast carcinoma, canine lung carcinoma, canine lymphoma, canine cutaneous mast cell tumor, human pancreatic/gastrointestinal neuroendocrine tumor, and unlabeled human melanoma. The final test set contained 10 independent tumor domains, introduced a new species: feline, and exposed models to new scanner/laboratory combinations and spindle-cell morphology. This was the edition in which MIDOG ceased to be primarily a scanner challenge and became a multi-axis domain-generalization challenge (Aubreville et al., 2023).

MIDOG 2025 is presented as the third edition of the series and as a substantial redesign rather than a simple continuation. The decisive change was moving beyond expert-selected hotspots to include random tissue ROIs and challenging ROIs enriched for hard negatives, artifacts, inflammation, and necrosis. The challenge also became explicitly two-track: Track 1 for mitotic figure detection and Track 2 for atypical mitotic figure classification (Bourgade et al., 29 Aug 2025, Aubreville et al., 5 Jun 2026).

3. Annotation protocols and evaluation criteria

MIDOG’s annotation protocols were designed around the observation that mitoses are sparse and easy to miss, and that expert disagreement is substantial. In MIDOG 2021, the organizers used three expert pathologists from different institutions and countries. One expert screened all ROIs and marked mitoses plus hard negatives; a RetinaNet was then trained to propose additional candidates; these candidates were re-reviewed in a class-blinded way; and disagreements were resolved by a third expert. The released labels included approximate 50 px square boxes around mitoses and non-mitotic hard negatives (Aubreville et al., 2022).

The primary detection metric in MIDOG 2021 was micro-averaged $F_1$, computed over all slides:

$$F_1 = \frac{2 \sum_k^N \mathrm{TP}_k}{2 \sum_k^N \mathrm{TP}_k + \sum_k^N \mathrm{FN}_k + \sum_k^N \mathrm{FP}_k}.$$

A predicted detection counted as a true positive if its centroid lay within $7.5\,\mu m$ of a ground-truth mitosis. Multiple detections of the same object were counted as false positives. The 2021 paper also reported 10,000 bootstrap repetitions with 5th and 95th percentiles for confidence intervals (Aubreville et al., 2022).

MIDOG 2022 added a second reference standard beyond the official three-expert consensus: a PHH3-assisted ground truth. The organizers de-stained H&E slides, re-stained them with phospho-histone H3 (PHH3), digitized the immunohistochemistry slides, aligned H&E and PHH3, and annotated with a superimposed view of both stains. This created an important benchmark-level distinction between official ranking labels and a post hoc reference intended to reduce reliance on subjective H&E-only judgments (Aubreville et al., 2023).

MIDOG 2025 continued PHH3-supported annotation for newly added detection ROIs and used the open-source EXACT annotation tool for co-registered H&E/PHH3 viewing. For Track 2, atypical versus normal labels were assigned by two pathologists, with disagreements resolved by a third expert. The paper reports Cohen’s $\kappa = 0.48$ for Track 1 and $\kappa = 0.68$ for Track 2. Track 1 evaluation used $F_1$ as the primary metric, with FROC-AUC and AP as secondary metrics; Track 2 used balanced accuracy and ROC AUC (Aubreville et al., 5 Jun 2026).

4. Detection paradigms in the series

The detection track has never converged to a single dominant architecture. The 2021 and 2022 overviews describe a heterogeneous field including single-stage detectors, multi-stage pipelines, semantic or instance segmentation, domain-adversarial training, stain normalization, style-transfer augmentation, Fourier-domain adaptation, and ensembling/test-time augmentation. In 2021, the winning method was AI medical with $F_1 = 0.748$ and the runner-up was TIA Centre with $F_1 = 0.764$0; the organizers specifically noted that all three best-performing methods included an auxiliary mitosis segmentation task, and that ensembling or test-time augmentation was common among top submissions (Aubreville et al., 2022). In 2022, the best overall performance came from TIA Centre with $F_1 = 0.764$1, followed closely by TCS Research with $F_1 = 0.764$2, despite the use of quite different pipelines (Aubreville et al., 2023).

Representative challenge reports illustrate the range of technical strategies. One MIDOG 2021 submission fused a DetectoRS detector with a deep ensemble classification model, using Macenko stain normalization, color-focused augmentation, and score fusion

$F_1 = 0.764$3

with optimal $F_1 = 0.764$4, and achieved $F_1 = 0.764$5 on the preliminary testing set (Liang et al., 2021). A MIDOG 2022 report showed that a single EfficientNet-B3 classifier could be converted into a detector using sliding-window inference and Grad-CAM++, with positive windows thresholded at 0.84, NMS threshold 0.22, and CAM hotspot centroids used as mitosis locations; this system reached $F_1 = 0.764$6 on the final test phase of task 1 (Gu et al., 2022).

Another 2022 report emphasized rotation invariance and fine-grained hard negative mining. It used a customized 70-layer ResNet with P4-group convolutional layers, dense whole-image inference, and a hard-negative retention search that ultimately trained the best model with only 19.6% of the overall pixel count of the training partition, while still obtaining about 0.697 on the final test and third place overall (Lafarge et al., 2023). A separate 2022 entry combined YOLOv5s with a Radial-Prediction Domain Adaption Classifier (RP-DAC) and HED-space stain augmentation, reaching $F_1 = 0.764$7 on the official test set (Annuscheit et al., 2022).

The 2025 detection literature broadened the architectural space further. A YOLOv12-m detector trained only on organizer-provided data reported $F_1 = 0.764$8, precision = 0.808, and recall = 0.794 on the preliminary test set, without external data or model ensembling (Bourgade et al., 29 Aug 2025). An RF-DETR submission, strengthened by hard negative mining from necrotic tissue, achieved $F_1 = 0.764$9, recall = 0.839, and precision = 0.746 on the preliminary test set (Giedziun et al., 29 Aug 2025). Team Westwood used nnUNetV2 for high-sensitivity candidate screening, followed by a random forest over three CNNs, and reported $F_1 = 0.740$0 on the preliminary test set (Xu et al., 29 Aug 2025). A separate two-stage study combined improved YOLO11x proposals with ConvNeXt-Tiny filtering and, on a fused dataset of MIDOG++, MITOS_WSI_CCMCT, and MITOS_WSI_CMC, achieved $F_1 = 0.740$1, which was 0.035 higher than the single-stage YOLO11x baseline (Xiao et al., 1 Sep 2025).

5. The introduction of atypical mitosis classification

MIDOG 2025 Track 2 formalized a new task: binary classification of mitotic crops into NMFs and AMFs. The official challenge overview presents this as a biologically motivated extension of the series and reports that a new public dataset of atypical versus normal mitoses was released for the track, built from 11,939 mitotic figures spanning 7 domains (Aubreville et al., 5 Jun 2026). Challenge reports then expanded this corpus with additional public sources such as AMi-Br, LUNG-MITO, GBM-TCGA, AtNorM-Br, and OMG-Octo Atypical, typically to increase domain diversity and AMF prevalence (Giedziun et al., 29 Aug 2025, Feki et al., 29 Aug 2025).

Methodologically, Track 2 has been dominated by foundation-model adaptation and strong augmentation. One solution used the pathology foundation model H-optimus-0 with LoRA / PEFT, soft labels from multi-expert consensus, hard negative mining, adaptive focal loss, supervised contrastive regularization, and a gradient-reversal domain head, reporting mean balanced accuracy $F_1 = 0.740$2 under 10-fold leave-one-domain-out cross-validation (Giedziun et al., 29 Aug 2025). Another strong baseline used DINOv3-H+ pretrained on natural images, with LoRA applied only to query and value projections and about 650k trainable parameters, reaching balanced accuracy $F_1 = 0.740$3 on the preliminary test set (Balezo et al., 28 Aug 2025).

Other competitive submissions explored lightweight regularization and augmentation recipes. The “Mix, Align, Distil” approach combined MixStyle, CBAM-based feature alignment using weak domain labels, and EMA-teacher distillation, attaining balanced accuracy $F_1 = 0.740$4, sensitivity $F_1 = 0.740$5, specificity $F_1 = 0.740$6, and ROC AUC $F_1 = 0.740$7 on the organizer-run preliminary leaderboard (Atey et al., 28 Aug 2025). A ConvNeXt Small model trained with a histopathology-specific augmentation pipeline and weighted sampling achieved balanced accuracy $F_1 = 0.740$8 and ROC AUC $F_1 = 0.740$9 on the preliminary leaderboard (Feki et al., 29 Aug 2025). A UNI2-h study found that Visual Prompt Tuning (VPT) outperformed a LoRA baseline, and that adding Vahadane and Macenko stain-normalization TTA yielded balanced accuracy $F_1$0 and ROC-AUC $F_1$1 (Meng et al., 1 Sep 2025).

Track 2 did not settle the question of whether pathology foundation models are always superior. One MIDOG 2025 report states that the best result apparently came not from a foundation model, but from a from-scratch ConvNeXt-Tiny classifier, which the authors speculate may reflect dataset size limitations (Shen et al., 28 Aug 2025). The official challenge overview reported a top Track 2 performance of balanced accuracy $F_1$2, with the top three teams separated by less than one percentage point (Aubreville et al., 5 Jun 2026). This suggests a benchmark in which strong solutions exist, but the margin between leading systems is narrow.

6. Recurrent findings, limitations, and scientific significance

Across editions, the series has produced several stable findings. First, domain generalization is possible but incomplete. MIDOG 2021 showed that scanner-agnostic detection could approach expert level, with the winning algorithm at $F_1$3 and the Top5 ensemble at 0.773 (Aubreville et al., 2022). MIDOG 2022 showed that performance remained strong in a broader setting—top $F_1$4—but also that new species, spindle-cell morphology, and a new scanner/lab combination caused small but significant decreases in performance (Aubreville et al., 2023). MIDOG 2025 then demonstrated that contextual variation inside the slide is itself a major source of failure: mean $F_1$5 was 0.735 in hotspot ROIs, 0.638 in random ROIs, and 0.479 in challenging ROIs, while precision dropped from 0.805 in hotspots to 0.400 in challenging regions and false detections increased by approximately 208% (Aubreville et al., 5 Jun 2026).

Second, the series repeatedly shows that annotation uncertainty is structurally important rather than incidental. In MIDOG 2022, re-evaluation against PHH3-assisted ground truth increased the total mitotic figure count by 15.0%, and 28.78% of PHH3-identified mitoses were not included in the original three-expert H&E majority vote. All methods then showed lower recall, although the ranking changed only minimally (Aubreville et al., 2023). A plausible implication is that MIDOG benchmarks measure not only generalization capacity but also robustness to epistemic uncertainty in the reference standard.

Third, ensembling is one of the few interventions that generalized across editions. The 2025 overview reports mean gains of 1.5 percentage points in $F_1$6 and 1.3 percentage points in balanced accuracy from ensembling, whereas TTA showed no relevant improvement (Aubreville et al., 5 Jun 2026). That finding is consistent with earlier editions, where many of the strongest methods used either explicit ensembles or multi-stage consensus mechanisms (Aubreville et al., 2022).

Several failure modes recur across challenge reports. One retrospective 2025 report documents a system with recall 0.9528 but precision 0.1267 on its best submission, concluding that the main problem was false positives from morphologically similar imposters across tissue domains; in the same study, adding a classifier ensemble to the detector reduced the full-pipeline $F_1$7 on PooledDB to 0.0646 because many true positives were also removed (Song et al., 1 Sep 2025). Another 2025 report isolates a telophase double-detection problem, in which one telophase event is predicted as two daughter-nucleus boxes, and resolves it with a distance-based aggregation step merging centroids within 10 $F_1$8 (Shen et al., 28 Aug 2025). The CAM-based 2022 detector likewise notes a limitation when multiple mitoses fall inside one positive patch: the CAM may strongly highlight only one and underrepresent the others (Gu et al., 2022).

The broader significance of the MIDOG Competition Series is therefore methodological and epistemic as much as leaderboard-oriented. It established scanner shift as a benchmarkable problem, expanded domain generalization to tumor type, species, and laboratory variation, and then exposed a further blind spot by showing that hotspot performance does not reliably predict performance in random or imposter-rich tissue (Aubreville et al., 2022, Aubreville et al., 2023, Aubreville et al., 5 Jun 2026). This suggests that MIDOG has become a benchmark not merely for mitosis detection, but for real-world pathology robustness.