MIDOG++: A Multi-Domain Mitotic Figure Benchmark
- MIDOG++ is an extensive, heterogeneous dataset and benchmark for mitotic figure detection and atypical classification, designed to evaluate model robustness under varied domain shifts.
- It comprises 11,937 annotations from 503 cases across seven tumor types and nine domain variants, capturing diverse imaging conditions including different scanners, stains, and tissue contexts.
- The dataset underpins advanced research in domain generalization, prompting innovative methods like hard negative mining and ensemble techniques to balance high recall with false-positive suppression.
MIDOG++ is an extended multi-domain dataset and benchmark for mitotic figure analysis in histopathology, developed within the MIDOG challenge lineage to study domain generalization rather than only in-domain detection accuracy. In the 2025 challenge literature, it is repeatedly described as a “comprehensive multi-domain dataset for mitotic figure detection,” and, in the newer benchmark formulation, it also underpins atypical mitotic figure classification. Its distinguishing property is deliberate heterogeneity across tumor types, species, scanners, staining protocols, and acquisition conditions, so that models must remain reliable under realistic distribution shift rather than only on curated hotspot regions (Giedziun et al., 29 Aug 2025, Aubreville et al., 5 Jun 2026).
1. Historical position in the MIDOG program
The MIDOG series originated as a benchmark for scanner-agnostic mitosis detection. MIDOG 2021 focused on scanner-induced domain shift in breast cancer ROIs, using 200 training cases and an additional 100 test cases distributed across six scanners, with two test scanners unseen during training; the winning individual method achieved an score of $0.748$ and the top-5 ensemble reached $0.773$ (Aubreville et al., 2022). MIDOG 2022 broadened the setting from scanner shift to stronger variation in scanner, tissue type, and species, with hidden evaluation on ten unseen domains, and challenge solutions increasingly emphasized hard negative mining, rotation invariance, and cross-domain robustness rather than pure detector capacity (Lafarge et al., 2023).
MIDOG++ represents the expanded dataset regime that supports this transition from scanner robustness to broader biological and contextual generalization. In the 2025 literature it is described as the extended multi-domain dataset that underpins both ordinary mitosis detection and the new atypical mitotic figure track, and MIDOG 2025 is framed as the largest and most comprehensive mitosis-annotated setting in the series, extending evaluation beyond classical hotspots to arbitrary tumor tissue and difficult regions (Aubreville et al., 5 Jun 2026, Bourgade et al., 29 Aug 2025).
2. Dataset composition and reported dataset variants
Challenge papers describe MIDOG++ in several closely related forms, reflecting its use across detection and classification tasks.
| Source | Reported MIDOG++ description | Task context |
|---|---|---|
| (Giedziun et al., 29 Aug 2025) | 11,937 mitotic figure annotations; 503 cases; seven tumor types; nine domain variants | Track 1 training dataset |
| (Bourgade et al., 29 Aug 2025) | 503 manually selected ROIs; 0.25 µm/px; roughly 7000 × 5000 pixels; human and canine samples | Detection training |
| (Xiao et al., 1 Sep 2025) | 553 patches; 7200 × 5400 pixels; 503 tumor cases; seven tumor types; 11,937 manually annotated mitotic figures | Fused detection dataset |
| (Percannella et al., 28 Aug 2025) | 11,939 mitotic figures; seven domains | Atypical subclassification setting |
One 2025 Track 1 report characterizes MIDOG++ as containing 11,937 mitotic figure annotations across 503 cases from seven tumor types and spanning nine domain variants, with value derived from multiple tumor types, different species, and heterogeneous acquisition conditions (Giedziun et al., 29 Aug 2025). Another challenge report specifies the tumor inventory as human breast carcinoma, canine lung carcinoma, canine lymphosarcoma, canine cutaneous mast cell tumor, human neuroendocrine tumor, canine soft tissue sarcoma, and human melanoma, with 503 manually selected ROIs scanned at 0.25 µm/px and ROI size roughly 7000 × 5000 pixels (Bourgade et al., 29 Aug 2025).
For detection experiments that combine MIDOG++ with canine whole-slide datasets, one paper reports MIDOG++ as 553 patches of size 7200 × 5400 from 503 tumor cases and 11,937 manually annotated mitotic figures (Xiao et al., 1 Sep 2025). For Track 2, the MIDOG 2025 atypical training set is described as the atypical mitotic figure subclassification of the entire MIDOG++ dataset, containing 11,939 mitotic figures across seven domains (Percannella et al., 28 Aug 2025). A plausible implication is that these small discrepancies reflect task-specific exports, patching conventions, or closely related benchmark releases rather than a contradiction about the benchmark’s overall role.
3. Benchmark design: from hotspot detection to multi-context generalization
The central design principle of MIDOG++ is robustness under distribution shift. The task is difficult not only because mitotic figures are rare and visually confusable with apoptotic cells, compressed nuclei, debris, necrosis, and inflammatory structures, but also because model performance degrades across laboratories, scanners, stains, tissue types, and species (Giedziun et al., 29 Aug 2025).
MIDOG 2025 redefined what performance on MIDOG++ should mean by extending evaluation beyond standard tumor hotspots. The final test set comprised 365 ROIs from 122 patient cases or WSIs across 12 tumor domains, digitized on multiple scanning platforms and including human, canine, and feline material. Track 1 explicitly distinguished three ROI contexts: hotspot ROIs selected for high mitotic density, random tissue ROIs sampled from tissue area with at least 80% tissue coverage, and challenging ROIs deliberately enriched with mitosis imposters such as apoptotic or necrotic cells, hyperchromatic nuclei, inflammatory cells, ink, folds, and blur (Aubreville et al., 5 Jun 2026).
This multi-context design changes the interpretation of algorithmic success. In MIDOG 2025, mean Track 1 across participants was $0.735$ in hotspots, $0.638$ in random ROIs, and $0.479$ in challenging ROIs, while recall remained relatively stable and precision deteriorated sharply; the false detection rate in challenging regions increased by 208% relative to hotspots (Aubreville et al., 5 Jun 2026). The benchmark therefore measures not only whether a detector can find mitoses in favorable regions, but whether it can suppress false positives in ordinary and adversarial tissue contexts. That logic extends the earlier MIDOG evaluation framework, where object matching already used a tight spatial tolerance of 7.5 µm to emphasize clinically meaningful localization (Aubreville et al., 2022).
4. MIDOG++ as a training substrate for mitotic figure detection
MIDOG++ functions as the principal supervised training resource for many 2025 Track 1 systems. A representative example is the RF-DETR approach, which treats the task as domain-generalized mitotic figure detection and uses MIDOG++ as the sole training source for its best model. That submission employed RF-DETR-Large with EMA and achieved preliminary-test , recall , and precision , with per-domain $0.748$0 ranging from $0.748$1 to $0.748$2; the authors explicitly argue that MIDOG++ is valuable because its breadth across tumor types, scanners, staining protocols, and species makes it a strong foundation for domain-generalized detection, while also warning that performance is sensitive to domain balance and split composition (Giedziun et al., 29 Aug 2025).
Other systems use MIDOG++ together with canine datasets such as MITOS_WSI_CMC and MITOS_WSI_CCMCT. A YOLOv12-m detector trained on MIDOG++, CMC, and CCMCT, with enforced 50% human-derived and 50% canine-derived images per batch and Multi-target Macenko normalization, reported preliminary-test $0.748$3, precision $0.748$4, and recall $0.748$5 without external data (Bourgade et al., 29 Aug 2025). Another two-stage FCOS-based study found that adding CMC to MIDOG++ improved validation performance from $0.748$6 to $0.748$7, whereas adding CCMCT together with CMC reduced it to $0.748$8, illustrating that extra data do not automatically improve generalization when the additional domain shift is too strong or poorly aligned with the target benchmark (Xu et al., 29 Aug 2025).
MIDOG++ has also exposed the limits of straightforward two-stage false-positive suppression. A Faster R-CNN–based challenge report achieved its best submission with a detector trained solely on MIDOG++, but only at $0.748$9, with recall $0.773$0 and precision $0.773$1, leading the authors to conclude that candidate generation is easier than robust false-positive reduction under domain shift (Song et al., 1 Sep 2025). Likewise, an FCOS system augmented with a FAL-CNN verifier and fusion network achieved $0.773$2, below the reported FCOS baseline of $0.773$3, showing that added architectural complexity can degrade rather than improve MIDOG++ performance when detector and verifier are mismatched (Broad et al., 29 Aug 2025).
Methodologically, MIDOG++ has encouraged a recurring set of design patterns: recall-first proposal generation, explicit hard negative mining, stain-aware augmentation, and domain balancing. Yet the benchmark has also shown that common normalization or augmentation heuristics are not uniformly beneficial; one 2025 report states that Macenko, Multi-Macenko, Reinhard, hematoxylin-only or eosin-only normalization, CutMix, and Gaussian blur all performed worse than training on unmodified MIDOG++ data (Giedziun et al., 29 Aug 2025). This aligns with earlier MIDOG 2022 evidence that careful fine-grained hard negative mining could produce strong generalization even when using only 19.6% of the training partition’s pixel count, rather than indiscriminately using all available negatives (Lafarge et al., 2023).
5. MIDOG++ and atypical mitosis classification
MIDOG++ is not limited to ordinary mitosis detection. In MIDOG 2025, all confirmed mitotic figures from hotspot regions were re-reviewed to form a public atypical mitotic figure subset, creating a Track 2 benchmark of 11,939 mitotic figures across seven domains, centered in 128 × 128 patches and labeled as normal or atypical (Aubreville et al., 5 Jun 2026). One challenge paper describes the MIDOG 2025 atypical training set as the atypical subclassification of the full MIDOG++ dataset and reports a leave-one-domain-out multi-task model achieving balanced accuracy $0.773$4 on the preliminary MIDOG25 Track 2 test set (Percannella et al., 28 Aug 2025).
This Track 2 extension has stimulated a distinct methodological literature. A teacher-student segmentation-centered model with contrastive learning and domain-adversarial training reported preliminary balanced accuracy $0.773$5 while sharing features between detection and atypical classification (Choe et al., 3 Sep 2025). An EfficientViT-L2 study combined MIDOG++ and AMi-Br into a unified dataset of 13,938 nuclei from seven cancer types and obtained preliminary balanced accuracy $0.773$6, ROC AUC $0.773$7, and raw accuracy $0.773$8 (Qi et al., 28 Aug 2025). A pathology foundation model ensemble using LoRA adaptation, fisheye transforms, and Fourier Domain Adaptation achieved overall preliminary balanced accuracy $0.773$9 (Ochi et al., 29 Aug 2025). A lighter training-time recipe based on MixStyle, weak-domain-label alignment, and EMA-teacher distillation reported balanced accuracy 0, sensitivity 1, specificity 2, and ROC AUC 3 on the organizer-run preliminary leaderboard (Atey et al., 28 Aug 2025).
The benchmark-level picture is that atypical mitosis classification is more consistent than ordinary detection but still domain-dependent. MIDOG 2025 reports Track 2 balanced accuracy values up to 4, with significant differences across domains and Cohen’s 5 for Track 2 annotation agreement, compared with 6 for Track 1 (Aubreville et al., 5 Jun 2026). This suggests that MIDOG++ has expanded from a detector benchmark into a broader platform for studying rare-event morphology, expert disagreement, and subtype recognition under domain shift.
6. Scientific significance, common misconceptions, and open issues
MIDOG++ matters because it operationalizes a stronger notion of domain generalization than earlier scanner-only benchmarks. It is not simply a larger MIDOG 2021 dataset. Rather, it combines biological diversity, species variation, scanner variability, staining variability, and contextual difficulty, and later benchmark formulations show that hotspot-only evaluation can substantially overestimate real-world reliability (Aubreville et al., 5 Jun 2026).
A common misconception is that more heterogeneous data necessarily solve generalization. MIDOG++-based studies argue the opposite: heterogeneity is the source of both value and difficulty. One challenge paper notes that identical architectures and hyperparameters can produce lower scores under different MIDOG++ train/validation splits, underscoring sensitivity to domain composition; the same paper suggests that exposure to diverse tissue types may matter more than concentration on a single cancer type (Giedziun et al., 29 Aug 2025). Another detection study shows that adding one external canine dataset to MIDOG++ helps, whereas adding a second one can hurt, indicating that dataset breadth must be balanced against domain compatibility (Xu et al., 29 Aug 2025).
A second misconception is that the principal failure mode is missed mitoses. In the MIDOG 2025 analysis, recall was relatively stable across hotspot, random, and challenging ROIs, whereas precision collapsed in the hard regions. The dominant unresolved problem is therefore false-positive control in necrotic, inflamed, artifact-rich, or otherwise deceptive tissue (Aubreville et al., 5 Jun 2026). This is consistent with challenge submissions that achieved very high recall but unusably low precision when transferred across domains (Song et al., 1 Sep 2025).
The benchmark also clarifies the practical value of ensemble methods. Post-challenge analysis reports mean gains of 1.5 percentage points in Track 1 7 and 1.3 percentage points in Track 2 balanced accuracy from ensembling, whereas test-time augmentation showed no relevant improvement (Aubreville et al., 5 Jun 2026). This suggests that model diversity currently contributes more to MIDOG++ robustness than repeated inference over augmented views.
Overall, MIDOG++ has become a reference benchmark for studying mitosis analysis “in the wild.” Its main scientific contribution is not only scale, but a task definition in which scanner shift, species shift, tumor shift, and contextual shift are treated as first-class variables. The resulting literature shows that MIDOG++ is simultaneously a productive training resource and a stringent stress test: strong models can achieve competitive performance, but false positives, split sensitivity, and context-dependent degradation remain central unsolved problems (Giedziun et al., 29 Aug 2025, Aubreville et al., 5 Jun 2026).