MIDOG 2025: Mitosis Domain Generalization Benchmark
- MIDOG 2025 Challenge is a domain generalization benchmark that evaluates mitosis detection and classification across varied tumor types, species, and tissue regions.
- It comprises two tracks focusing on object-level detection in WSIs and binary classification distinguishing normal from atypical mitoses using rigorous multi-expert annotations.
- Performance metrics, including an F1 score up to 0.740 and balanced accuracy up to 0.908, highlight both methodological advances and challenges in real-world histopathology.
Searching arXiv for recent MIDOG 2025 challenge papers and the main overview paper. The Mitosis Domain Generalization (MIDOG) 2025 Challenge is the third large benchmark in the MIDOG series and frames mitosis analysis as a problem of domain generalization under realistic histopathology variability. It was organized as a MICCAI 2025 satellite event and consists of two tracks: Track 1, detection of mitotic figures as objects in histology images, and Track 2, classification of confirmed mitotic figures into normal versus atypical mitotic figures (AMFs) (Aubreville et al., 5 Jun 2026). MIDOG 2025 extends earlier MIDOG editions by evaluating models not only across multiple tumor types, species, and scanners, but also across distinct tissue contexts rather than hotspot-only regions, thereby shifting the benchmark toward what the organizers describe as detection “in the wild” (Aubreville et al., 5 Jun 2026). In the challenge overview, the test dataset comprises 365 cases, spanning 12 distinct human, canine and feline tumor types, with 18 teams in the detection track achieving F1 scores ranging up to 0.740, and 21 submissions in the AMF classification track achieving balanced accuracy values up to 0.908 (Aubreville et al., 5 Jun 2026).
1. Challenge scope and conceptual framing
MIDOG 2025 was designed to test whether mitosis analysis algorithms can operate robustly across many tumor types, species, scanners, and tissue regions, not merely across scanner-induced variation or pathologist-selected hotspots (Aubreville et al., 5 Jun 2026). This distinguishes it from earlier MIDOG editions: MIDOG 2021 focused on scanner-induced domain shift, while MIDOG 2022 extended to multiple tumor types and species but still restricted evaluation to expert-selected hotspot regions (Aubreville et al., 5 Jun 2026). A central premise of MIDOG 2025 is that clinical deployment requires robustness in the broader histological landscape, including random tissue areas and regions rich in hard negatives (Aubreville et al., 5 Jun 2026).
The challenge is motivated by the clinical importance of mitotic count (MC) and mitotic characterization in tumor grading and prognostication. Manual assessment is described as being affected by inter-rater agreement limitations and sampling bias, especially during hotspot selection (Aubreville et al., 5 Jun 2026). MIDOG 2025 therefore asks not only whether a model can find mitotic figures, but also whether it can distinguish normal from atypical mitotic figures, since AMFs are described as emerging prognostic markers (Aubreville et al., 5 Jun 2026).
The formal distinction between the two tracks is stable across the challenge literature. Track 1 addresses mitotic figure detection under strong domain shift (Bourgade et al., 29 Aug 2025), whereas Track 2 addresses binary classification of mitotic figure crops into normal mitotic figures (NMFs) and atypical mitotic figures (AMFs) (Yamagishi et al., 26 Aug 2025), or equivalently typical versus atypical mitoses (Percannella et al., 28 Aug 2025). A plausible implication is that MIDOG 2025 links two historically separate tasks—object-level mitosis detection and subtype-level mitosis interpretation—within a single generalization benchmark.
2. Test set design, domains, and annotation protocol
For Track 1, the challenge overview states that the test set contains 122 whole slide images (WSIs), each from a different patient, covering 12 tumor domains and multiple scanners (Aubreville et al., 5 Jun 2026). For each WSI, the organizers selected three non-overlapping 2 mm² regions of interest (ROIs): a hotspot ROI, a random ROI, and a challenging ROI (Aubreville et al., 5 Jun 2026). One WSI was too small for non-overlapping random selection, yielding a final total of 365 ROIs: 122 hotspots, 121 random, and 122 challenging (Aubreville et al., 5 Jun 2026). The hotspot ROI represents the traditional high-proliferation region; the random ROI is sampled uniformly from tissue regions with at least 80% tissue coverage; the challenging ROI is selected for high densities of imposters and artifacts such as apoptotic or necrotic cells, hyperchromatic nuclei, inflammatory cells, ink markings, out-of-focus artifacts, tissue folds, and delayed fixation areas (Aubreville et al., 5 Jun 2026).
For Track 2, the challenge overview states that the dataset is derived from MIDOG++, with 11,939 mitotic figures across seven tumor domains, represented as 128×128 px patches centered on mitotic figures (Aubreville et al., 5 Jun 2026). Several Track 2 solution papers describe the official training data in compatible terms. One paper reports 10,191 NMFs, 1,748 AMFs, 454 labeled images, and 9 distinct domains in the official MIDOG 2025 Track 2 dataset (Ramchandani et al., 21 Sep 2025). Another reports the official training dataset as containing 10,191 instances of NMFs, 1,748 instances of AMFs, 11,939 total mitotic annotations, and 454 labeled images, with a test set of 120 cases covering 12 tumor types from human and veterinary pathology (Yamagishi et al., 26 Aug 2025). A plausible implication is that different participating teams describe closely related but not always identically enumerated domain partitions of the Track 2 material.
The annotation workflow for Track 1 used a PHH3-assisted, multi-expert workflow (Aubreville et al., 5 Jun 2026). The H&E slide was scanned, the coverslip removed, the H&E stain washed out, and PHH3 immunohistochemistry applied to the same section; H&E and PHH3 WSIs were then registered using an affine transform (Aubreville et al., 5 Jun 2026). In the first annotation phase, a pathologist classified each candidate into five categories combining PHH3 status and morphology; classes 1–3 were treated as mitotic figures, classes 4–5 as hard negatives (Aubreville et al., 5 Jun 2026). A second pathologist then reviewed 30 μm H&E patches in a blinded phase, with a third expert as tie-breaker in cases of disagreement (Aubreville et al., 5 Jun 2026). The challenge overview reports Cohen’s for Track 1 and Cohen’s for Track 2 (Aubreville et al., 5 Jun 2026).
For Track 2, AMF versus NMF labels were assigned on confirmed MF patches, with two pathologists independently labeling each mitosis according to the morphological criteria summarized from Donovan et al. and a third expert resolving disagreements (Aubreville et al., 5 Jun 2026). AMFs are described as mitoses with abnormal division patterns such as multipolar spindles, lagging chromosomes, highly irregular chromatin, or related abnormalities (Aubreville et al., 5 Jun 2026, Kotte et al., 1 Sep 2025).
3. Evaluation framework and challenge outcomes
For Track 1, the primary metric is the micro-averaged F1 score (Choe et al., 3 Sep 2025, Aubreville et al., 5 Jun 2026). One Track 1 paper explicitly states that MIDOG 2025 evaluates detection with a micro F1-score, where a detection counts as a true positive if its centroid lies within of a ground-truth mitosis (Choe et al., 3 Sep 2025). The challenge overview further reports that performance was analyzed across hotspot, random, and challenging ROIs, and that although most models performed reliably in traditional hotspots, there was significant performance degradation in challenging ROIs, where false positive rates tripled (Aubreville et al., 5 Jun 2026). It also states that performance varied significantly across the 12 tumor types, exposing “blind spots” in current architectures for rare or pleomorphic malignancies (Aubreville et al., 5 Jun 2026).
For Track 2, the primary metric is balanced accuracy (Percannella et al., 28 Aug 2025, Yamagishi et al., 26 Aug 2025, Ramchandani et al., 21 Sep 2025). Several participating papers reproduce the binary definition,
or equivalent formulations in terms of sensitivity and specificity (Yamagishi et al., 26 Aug 2025, Ramchandani et al., 21 Sep 2025). This metric is repeatedly justified as appropriate for the substantial AMF/NMF class imbalance (Percannella et al., 28 Aug 2025, Yamagishi et al., 26 Aug 2025). The challenge overview reports that Track 2 submissions achieved balanced accuracy values up to 0.908 (Aubreville et al., 5 Jun 2026).
The organizers also evaluated the effect of common inference-time strategies. According to the challenge overview, ensembling yielded mean increases of 1.5 percentage points in F1 score for Track 1 and 1.3 percentage points in balanced accuracy for Track 2, whereas TTA showed no relevant improvement (Aubreville et al., 5 Jun 2026). This is noteworthy because many participating methods nevertheless relied on either cross-validation ensembles or explicit TTA, suggesting that empirical utility was method-dependent even though the aggregate challenge-level analysis favored ensembling over TTA (Aubreville et al., 5 Jun 2026).
A compact summary of the official challenge-level outcomes is given below.
| Component | Reported value | Source |
|---|---|---|
| Track 1 participating teams | 18 | (Aubreville et al., 5 Jun 2026) |
| Track 1 top F1 | 0.740 | (Aubreville et al., 5 Jun 2026) |
| Track 2 submissions | 21 | (Aubreville et al., 5 Jun 2026) |
| Track 2 top balanced accuracy | 0.908 | (Aubreville et al., 5 Jun 2026) |
4. Methodological landscape in Track 1: detection under multi-context domain shift
The Track 1 ecosystem was dominated by object detection frameworks, particularly the YOLO family, although segmentation-based and two-stage systems were also represented (Aubreville et al., 5 Jun 2026). The challenge overview states that among analyzed Track 1 submissions, 10/14 used object detection frameworks and 4/14 used segmentation-based approaches (Aubreville et al., 5 Jun 2026). Supporting participant papers concretize this diversity.
One YOLO-based paper used YOLOv12-m as a single-stage detector configured to detect two classes, true mitosis and hard negatives, trained on MIDOG++, CMC, and CCMCT, and reported F1-score = 0.801, precision = 0.808, and recall = 0.794 on the preliminary MIDOG 2025 test set (Bourgade et al., 29 Aug 2025). That method combined multi-target Macenko stain normalization, human/canine-balanced sampling, background tiles, NMS, test-time augmentation, and Weighted Boxes Fusion (WBF) (Bourgade et al., 29 Aug 2025). Another Track 1 submission, based on RF-DETR, reported F1 = 0.789, recall = 0.839, and precision = 0.746 on the preliminary test set and highlighted the importance of domain-balanced training and hard negative mining while finding that several stain normalization strategies lowered F1 (Giedziun et al., 29 Aug 2025).
Two-stage designs were also common. One paper proposed a two-stage detection-classification framework using improved YOLO11x proposals followed by a ConvNeXt-Tiny classifier and reported F1-score = 0.882, compared with 0.847 for improved YOLO11x alone and 0.827 for a basic YOLO11x baseline, with precision increasing from 0.762 to 0.839 and recall remaining comparable (Xiao et al., 1 Sep 2025). Another paper used FCOS for candidate localization and a ResNet50 + EfficientNetB2 ensemble for candidate refinement, increasing F1 on MIDOG++ validation from 0.803 for the base detector to 0.8432 with classification refinement (Xu et al., 29 Aug 2025). By contrast, a cautionary report on a Faster R-CNN plus three-classifier pipeline achieved F1-score 0.2237, Recall 0.9528, and Precision 0.1267 on the challenge test set, and concluded that false-positive suppression remained a fundamental difficulty under diverse domains (Song et al., 1 Sep 2025).
Segmentation-centered approaches likewise appeared. A teacher–student UNet system formulated detection as pixel-level segmentation, incorporated contrastive representation learning and domain-adversarial training, and reported F1 = 0.7660 in Track 1 on the preliminary test set (Choe et al., 3 Sep 2025). A unified pipeline termed MitoDetect++ used an EfficientNetV2-L encoder within a U-Net-like detector and reserved Track 2 for a separate Virchow2-based classifier (Nasir et al., 28 Aug 2025). Team Westwood used nnUNetV2 for candidate screening followed by a random-forest ensemble of CNN classifiers and obtained F1 = 0.7450 on the preliminary Track 1 test set (Xu et al., 29 Aug 2025).
Taken together, these papers show that Track 1 was not merely a detector benchmark in the narrow sense. It was a benchmark for how to combine multi-domain training data, hard negative handling, stain-aware augmentation, tiling/fusion policies, and occasionally second-stage verification under a three-context ROI design that strongly penalized background sensitivity (Bourgade et al., 29 Aug 2025, Xiao et al., 1 Sep 2025, Giedziun et al., 29 Aug 2025, Aubreville et al., 5 Jun 2026).
5. Methodological landscape in Track 2: atypical mitosis classification
Track 2 attracted a particularly broad range of classifiers, from conventional CNNs to pathology-specific foundation models. The challenge overview states that most teams fine-tuned or adapted pre-trained models, including LoRA on vision foundation models such as UNI, Virchow, Virchow2, HIBOU, and DINOv3, as well as conventional fine-tuning of ConvNeXt, EfficientNet, ResNet, DenseNet, and Hover-Net (Aubreville et al., 5 Jun 2026).
A recurring design theme was object-centric restriction of context. A ConvNeXt V2 submission used 60% center cropping, 5-fold cross-validation, and a ConvNeXt V2 base model, and reported balanced accuracy = 0.8831 overall on the development phase, with an ablation showing that center cropping improved balanced accuracy from 0.8090 ± 0.0361 to 0.8492 ± 0.0065 and prevented non-convergent folds (Yamagishi et al., 26 Aug 2025). Another Track 2 paper using EfficientViT-L2 employed leave-one-cancer-type-out cross-validation with 5-fold ensembles, reporting balanced accuracy of 0.859, ROC AUC of 0.942, and raw accuracy of 0.85 in the preliminary evaluation phase (Qi et al., 28 Aug 2025).
A second major theme was multi-task or hardness-aware supervision. One submission introduced a multi-task neural network with a Pre-Activated ResNet-50 backbone, a classification head, a mitosis binary segmentation decoder, and a pixel-level mitosis classification decoder, trained with
and achieved balanced accuracy = 0.856 on the preliminary MIDOG25 Track 2 test set (Percannella et al., 28 Aug 2025). Another used a ResNet-50 backbone with expert-specific phenotype heads and a hardness head to jointly model phenotype and annotation difficulty, reaching balanced accuracy in 5-fold cross-validation and overall preliminary BA (Kotte et al., 1 Sep 2025).
Foundation-model approaches formed a third major cluster. A study of Virchow, Virchow2, and UNI with LoRA found that Virchow with LoRA rank 8 and a 3-fold ensemble achieved balanced accuracy of 88.37% on the preliminary test set and joint 9th place on the leaderboard (Ramchandani et al., 21 Sep 2025). Another submission based on UNI2-h reported that visual prompt tuning (VPT) substantially outperformed LoRA, and that adding Vahadane and Macenko stain normalization within TTA produced balanced accuracy = 0.8837 and ROC-AUC = 0.9513, ranking within the top 10 teams (Meng et al., 1 Sep 2025). A further foundation-model paper selected H-optimus-0 with LoRA, soft labels based on multi-expert consensus, adaptive focal loss, metric learning, hard negative mining, and domain adaptation, achieving mean BA under leave-one-domain-out evaluation (Giedziun et al., 29 Aug 2025).
More conventional but highly effective CNNs also remained competitive. A ConvNeXt Small model trained on all available datasets with a histopathology-specific augmentation pipeline achieved balanced accuracy = 0.8961 and ROC AUC = 0.9561 on the preliminary leaderboard (Feki et al., 29 Aug 2025). A DenseNet-121 framework with Macenko stain-aware augmentation, weighted sampling, and a hybrid objective combining class-weighted BCE and focal loss reported balanced accuracy 85.0%, AUROC 0.927, sensitivity 89.2%, and specificity 80.9% on the official test set (Dukre et al., 26 Oct 2025). A data-centric UCL submission, grounded in newly released datasets and modern training recipes, reported Track-2 balanced accuracy of 0.9107 for atypical mitotic cell classification on its own test set (Shen et al., 28 Aug 2025). This suggests that the challenge-level best value of 0.908 (Aubreville et al., 5 Jun 2026) coexisted with participant-side internal evaluations that occasionally exceeded that level on non-identical test partitions.
6. Interpretive significance and recurrent lessons
MIDOG 2025 formalized a transition from scanner-centric domain shift benchmarks to multi-context, multi-tumor, multi-species robustness testing. The defining insight of the challenge overview is that models that appear reliable in hotspots can degrade substantially in challenging ROIs, with false positive rates tripling (Aubreville et al., 5 Jun 2026). This suggests that the benchmark did not merely test morphology classification in isolation; it tested whether models could resist background-driven or context-driven failure modes in non-curated tissue regions.
Across both tracks, a consistent methodological pattern emerges. High-performing systems generally combined several of the following: broad domain exposure, histopathology-specific augmentation, object-centric cropping or localization, class-imbalance-aware objectives, and ensembling (Yamagishi et al., 26 Aug 2025, Percannella et al., 28 Aug 2025, Bourgade et al., 29 Aug 2025, Feki et al., 29 Aug 2025, Aubreville et al., 5 Jun 2026). By contrast, the aggregate challenge analysis found no relevant improvement from TTA (Aubreville et al., 5 Jun 2026), even though multiple strong participant systems used it (Bourgade et al., 29 Aug 2025, Nasir et al., 28 Aug 2025, Meng et al., 1 Sep 2025). This is best read not as a contradiction but as evidence that TTA was not a universally reliable lever once averaged over heterogeneous methods and domains.
Another recurring lesson concerns the tension between foundation models and specialized task structure. Several Track 2 papers show strong results with pathology-specific foundation models adapted by LoRA or VPT (Ramchandani et al., 21 Sep 2025, Giedziun et al., 29 Aug 2025, Meng et al., 1 Sep 2025), yet other papers report very competitive or stronger performance from carefully tuned CNNs such as ConvNeXt, DenseNet-121, or ResNet-50 (Feki et al., 29 Aug 2025, Kotte et al., 1 Sep 2025, Dukre et al., 26 Oct 2025). This suggests that in MIDOG 2025, large pretrained backbones were advantageous but not sufficient; architectural choice interacted strongly with sampling policy, augmentation strength, thresholding, and the way domain shift was operationalized.
A common misconception is that domain generalization in mitosis analysis is principally a problem of stain normalization. The participant literature gives a more qualified picture. Some methods benefited from Macenko or Vahadane-based stain augmentation or normalization (Bourgade et al., 29 Aug 2025, Meng et al., 1 Sep 2025, Dukre et al., 26 Oct 2025), whereas the RF-DETR paper explicitly reports that Macenko, Multi-Macenko, Reinhard, hematoxylin-only, eosin-only, CutMix, and Gaussian blur all decreased F1 in that setting (Giedziun et al., 29 Aug 2025). This suggests that stain handling is task- and architecture-dependent and that preserving native histomorphological variability can be as important as harmonizing color statistics.
In encyclopedia terms, MIDOG 2025 marks a methodological inflection point. It reframed mitosis analysis from a hotspot-optimized detection problem into a benchmark of clinical reliability under contextual diversity, while simultaneously elevating atypical mitosis classification into a first-class challenge task (Aubreville et al., 5 Jun 2026). The resulting body of work shows that the field can now assemble strong multi-domain systems, but also that “in the wild” mitosis analysis remains limited by rare morphologies, hard negatives, and cross-domain blind spots (Aubreville et al., 5 Jun 2026).