MIDOG 2025 Track 2

• MIDOG 2025 Track 2 is a domain-generalization benchmark that classifies mitotic figure image patches as normal or atypical, emphasizing robustness under varied tumor types, species, and staining conditions.
• The challenge leverages diverse methodologies including CNNs, vision transformers, and LoRA-enhanced foundation models, combined with data-centric strategies like augmentation and ensembling to manage domain shifts.
• Performance is primarily measured by balanced accuracy, revealing significant variability across tumor types and highlighting the importance of robust domain adaptation and ensemble techniques.

MIDOG 2025 Challenge Track 2 is a domain-generalization benchmark for binary classification of mitotic-figure image patches centered on confirmed mitoses into normal/typical versus atypical classes. In the official challenge formulation, Track 2 evaluates image patches extracted from hotspot regions and asks whether each mitotic figure is “normal” or “atypical,” with performance summarized primarily by balanced accuracy under substantial variation in tumor type, species, scanner, and staining conditions (Aubreville et al., 5 Jun 2026). Team reports describe the same task using closely related terminology such as “typical vs. atypical” or “NMF vs. AMF,” but they converge on the same technical problem: robust atypical-mitosis recognition under domain shift (Percannella et al., 28 Aug 2025, Atey et al., 28 Aug 2025).

1. Task definition and benchmark scope

Track 2 was introduced as the atypical mitotic figure classification component of MIDOG 2025. The official challenge paper defines it as binary classification of image patches centered on confirmed mitotic figures, with the label space consisting of normal versus atypical mitoses (Aubreville et al., 5 Jun 2026). A closely aligned description appears in Percannella et al., where the task is phrased as deciding whether a small crop of a histopathology image centered on a mitotic figure is a “typical” or “atypical” mitosis (Percannella et al., 28 Aug 2025).

The benchmark is explicitly designed around domain shift. Across the participating papers, domain variability is repeatedly attributed to scanner, stain, tissue type, species, laboratory, and acquisition differences (Percannella et al., 28 Aug 2025, Atey et al., 28 Aug 2025, Aubreville et al., 5 Jun 2026). The official challenge analysis further emphasizes that MIDOG 2025 moved toward a more realistic “in the wild” setting, and that Track 2 performance varies significantly across tumor types, revealing domain-specific “blind spots” in current architectures (Aubreville et al., 5 Jun 2026).

A preliminary evaluation phase and a final official evaluation both appear in the literature. Ochi and Bae report that the preliminary evaluation phase used four held-out domains, denoted domain_0 through domain_3, scored by balanced accuracy (Ochi et al., 29 Aug 2025). The official challenge paper reports pooled final-test results across 12 tumor types spanning human, canine, and feline cases (Aubreville et al., 5 Jun 2026). This suggests that leaderboard values reported in individual team papers and final challenge rankings should be interpreted as results from different evaluation phases rather than as directly interchangeable figures.

2. Dataset construction, patch extraction, and annotation

The final Track 2 test set described by Aubreville et al. was built from 122 patients, one whole-slide image per patient, spanning 12 tumor types: human melanoma, human astrocytoma, human bladder carcinoma, human colon carcinoma, human glioblastoma, human lung adenocarcinoma, human meningioma, canine mammary carcinoma, canine cutaneous mast cell tumor, canine hemangiosarcoma, feline soft-tissue sarcoma, and feline gastrointestinal lymphoma (Aubreville et al., 5 Jun 2026). Only hotspot ROIs of 2 mm² were used for Track 2 patch extraction, and for each hotspot ROI all ground-truth mitotic figures were cropped as square patches of approximately 30 µm side length, approximately 128 × 128 pixels (Aubreville et al., 5 Jun 2026).

The annotation protocol was multi-stage. A board-certified veterinary pathologist annotated all candidate mitotic figures in the hotspot ROI using PHH3-assisted registration between H&E and PHH3 IHC; two independent experts then classified every mitotic-figure patch as “normal” or “atypical”; discordant labels were reviewed by a third board-certified pathologist, who rendered the final decision (Aubreville et al., 5 Jun 2026). The reported inter-observer agreement between the two primary raters was Cohen’s $\kappa = 0.68$, described as substantial agreement (Aubreville et al., 5 Jun 2026).

Training data in participant reports were broader than the official test benchmark and often combined multiple public resources. Percannella et al. used the MIDOG 2025 Atypical Training Set containing 11,939 mitotic figures drawn from seven domains of the MIDOG++ corpus, with each domain defined as a distinct combination of scanner, center, tumor type, and species (Percannella et al., 28 Aug 2025). Several teams supplemented this set with AMi-Br, OMG-Octo Atypical, AtNorM-Br, AtNorM-MD, Shen et al. datasets, or other public sources to enlarge domain coverage (Feki et al., 29 Aug 2025, Balezo et al., 28 Aug 2025, Nasir et al., 28 Aug 2025, Giedziun et al., 29 Aug 2025). The prevalence of atypical mitoses was consistently low. For example, the official MIDOG 2025 training set summarized by Balezo et al. contained 11,939 crops, of which 1,748 were atypical (Balezo et al., 28 Aug 2025), while Kotte et al. reported the same 11,939 total patches with 10,191 NMF and 1,748 AMF (Kotte et al., 1 Sep 2025).

3. Evaluation protocol and reported performance

The official primary metric for Track 2 is balanced accuracy. In the challenge paper it is defined as

$$\mathrm{BA}=\tfrac12\left(\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}+\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}\right)$$

with ROC AUC used as a secondary metric (Aubreville et al., 5 Jun 2026). Percannella et al. explicitly state that balanced accuracy is the official Track 2 metric (Percannella et al., 28 Aug 2025).

On the official final test set, the challenge paper reports the following top-five Track 2 results (Aubreville et al., 5 Jun 2026):

Rank / Team	BA	ROC AUC
1. Balezo et al.	0.908	0.970
2. Nasir et al.	0.901	0.967
3. Yamagishi et al.	0.900	0.971
4. Qi et al.	0.897	0.962
5. Percannella et al.	0.897	0.962

The same paper reports that performance differed significantly across the 12 tumor types, with feline gastrointestinal lymphoma identified as the hardest domain and human astrocytoma as the easiest (Aubreville et al., 5 Jun 2026). Rare, highly pleomorphic tumors such as human glioblastoma and feline gastrointestinal lymphoma showed an average drop of 8 percentage points in atypical-class recall relative to the easiest domains (Aubreville et al., 5 Jun 2026).

Individual team papers provide complementary results on preliminary leaderboards, held-out internal splits, or external tests. Percannella et al.’s multi-task network obtained balanced accuracy 0.856 on the MIDOG25 Track 2 preliminary test set via an ensemble of fold-models with majority voting, alongside leave-one-domain-out mean test balanced accuracy 0.833 ± 0.051 and AMi-Br test balanced accuracy 0.806 ± 0.013 (Percannella et al., 28 Aug 2025). The “Mix, Align, Distil” method reported preliminary balanced accuracy 0.8762, sensitivity 0.8873, specificity 0.8651, and ROC AUC 0.9499 (Atey et al., 28 Aug 2025). A ConvNeXt small system trained on pooled datasets reported mean balanced accuracy 0.8961 and mean ROC AUC 0.9561 on the preliminary leaderboard (Feki et al., 29 Aug 2025). A LoRA-tuned DINOv3-H+ model reported preliminary balanced accuracy 0.8871 (Balezo et al., 28 Aug 2025). Shen et al. reported balanced accuracy 0.9107 on a held-out MIDOG 2025 atypical test set for their best ConvNeXt-tiny ensemble (Shen et al., 28 Aug 2025).

These figures are not directly comparable unless the evaluation phase is specified. The official challenge paper reports the final benchmark ceiling at 0.908 balanced accuracy across 12 tumor types (Aubreville et al., 5 Jun 2026), whereas many team papers report preliminary or internal evaluations on different splits (Atey et al., 28 Aug 2025, Feki et al., 29 Aug 2025, Balezo et al., 28 Aug 2025).

4. Architectural and algorithmic patterns

Track 2 submissions span CNNs, vision transformers, pathology foundation models, and multi-task hybrids. A recurring pattern is that domain robustness is pursued through augmentation, multi-domain training, parameter-efficient adaptation, ensembling, or auxiliary objectives rather than through a single dominant architecture.

A first cluster consists of convolutional classifiers and multi-task CNNs. Percannella et al. proposed a single convolutional neural network with a shared Pre-Activated ResNet-50 backbone, a main image-level classification head, and two U-Net-style auxiliary decoders for binary mitosis segmentation and pixel-wise atypicality prediction (Percannella et al., 28 Aug 2025). The total loss was

$$L_{\text{total}}=L_{\text{cls}}+L_{\text{seg}}+L_{\text{pix}},$$

with weighted binary cross-entropy for classification and Dice losses for the dense tasks (Percannella et al., 28 Aug 2025). A different CNN line is represented by ConvNeXt-based systems. One report used ConvNeXt small with a histopathology-specific augmentation pipeline, pooled AMi-Br, AtNorM-Br, MIDOG++, and OMG-Octo Atypical, and evaluated by grouped 5-fold cross-validation (Feki et al., 29 Aug 2025). Another used ConvNeXt V2 Base with 60% center cropping and a 5-fold ensemble, reporting that center cropping improved balanced accuracy from 0.8090 ± 0.0361 to 0.8492 ± 0.0065 on converged folds (Yamagishi et al., 26 Aug 2025). Shen et al. selected ConvNeXt-tiny as the best validation model among EfficientNet-B0/B3, ConvNeXt-tiny/small, and UNI (Shen et al., 28 Aug 2025).

A second cluster uses training-time regularizers targeted at domain shift. “Mix, Align, Distil” builds on DenseNet-121 with MixStyle layers inserted into early and mid dense blocks, CBAM refinement, an auxiliary alignment loss over weak domain labels, and knowledge distillation from an exponential moving average teacher (Atey et al., 28 Aug 2025). The combined training objective is

$$L_{\text{total}}=L_{\text{cls}}+\beta L_{\text{align}}+\gamma L_{\text{KD}},$$

with $\beta$ ramped by a DANN-style schedule and $\gamma$ warmed up to 0.5 (Atey et al., 28 Aug 2025). Kotte et al. took a different route by explicitly modeling both phenotype and “hardness” imbalance. Their ResNet-50 backbone feeds three phenotype heads corresponding to the three pathologist labels and one hardness head, all trained with focal loss and combined by

$$\mathcal{L}_{\mathrm{total}} = \theta \sum \mathcal{L}_{\mathrm{focal}}^{(\mathrm{phenotype})} + (1-\theta)\mathcal{L}_{\mathrm{focal}}^{(\mathrm{hardness})}, \qquad \theta=0.5$$

(Kotte et al., 1 Sep 2025).

A third cluster centers on foundation models and parameter-efficient fine-tuning. Balezo et al. adapted DINOv3-H+ using LoRA on query and value projections, with rank $r=4$, scaling $\alpha=8.0$, and approximately 650,000 trainable parameters (Balezo et al., 28 Aug 2025). MitoDetect++ used Virchow2 with LoRA inserted into qkv, output projection, and feed-forward layers, reducing trainable parameters to under 5 million while keeping the backbone frozen (Nasir et al., 28 Aug 2025). Ochi and Bae fine-tuned UNI, Virchow, and Virchow2 PFMs via LoRA, combined fisheye distortion and Fourier Domain Adaptation, and ensembled the models with learned soft-voting weights (Ochi et al., 29 Aug 2025). Meng et al. investigated UNI2-h with LoRA, then visual prompt tuning, and finally test-time augmentation plus Vahadane and Macenko stain normalization, reporting that VPT improved balanced accuracy from 0.8305 to 0.8711 and that VPT plus TTA reached 0.8837 on the preliminary leaderboard (Meng et al., 1 Sep 2025). Another foundation-model study used H-optimus-0 with LoRA, MixUp, soft labels derived from three pathologist votes, hard negative mining, adaptive focal loss, metric learning, and a gradient-reversal domain head (Giedziun et al., 29 Aug 2025).

A fourth pattern is data-centric reweighting. Sequential Hard Mining organizes training into up to three rounds, updating example weights according to misclassification and retraining with weighted sampling (Lafarge et al., 28 Aug 2025). The first hard-mining update raised internal balanced accuracy from 0.718 to 0.846 and preliminary test balanced accuracy to 0.812, while a third round did not improve validation performance (Lafarge et al., 28 Aug 2025).

5. Domain generalization strategies and challenge-level findings

The challenge literature converges on several domain-generalization strategies. Stain and color perturbation are nearly universal. Percannella et al. used random channel perturbations plus rotations and flips (Percannella et al., 28 Aug 2025). Bourgade et al., although focused on Track 1, describe multi-target Macenko stain transfer with 50 learned “stainers” as a domain-adaptation component (Bourgade et al., 29 Aug 2025). ConvNeXt-based Track 2 systems used extensive color, stain, blur, deformation, and noise augmentations (Feki et al., 29 Aug 2025), while Shen et al. combined RandAugment with H&E stain augmentation and TTA over flipped crops (Shen et al., 28 Aug 2025).

Balanced sampling and explicit class-imbalance treatment are equally common. Examples include inverse-frequency weighting in classification loss (Percannella et al., 28 Aug 2025), WeightedRandomSampler over class frequencies (Feki et al., 29 Aug 2025), focal loss with $\alpha=0.25$ and $$\mathrm{BA}=\tfrac12\left(\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}+\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}\right)$$0 (Nasir et al., 28 Aug 2025), and adaptive focal loss with a dynamic positive-class weight (Giedziun et al., 29 Aug 2025). Several teams also pooled multi-domain datasets to maximize exposure to inter-domain variability, often combining MIDOG 2025 with AMi-Br and OMG-Octo Atypical (Feki et al., 29 Aug 2025, Balezo et al., 28 Aug 2025, Shen et al., 28 Aug 2025).

At the challenge-analysis level, ensembling was the most consistently useful post-training intervention. The official paper reports that 12 of 20 teams ensembled multiple models, with mean ensemble size approximately 4.7, and that ensembling improved balanced accuracy by a mean of 1.299 percentage points and a median of 1.230 percentage points (Aubreville et al., 5 Jun 2026). In contrast, test-time augmentation produced a mean gain of only 0.299 percentage points and a median gain of 0.472 percentage points, with no team gaining more than 0.7 percentage points and two teams showing slight calibration degradation (Aubreville et al., 5 Jun 2026). This directly qualifies a common assumption that TTA is a major robustness lever in histopathology patch classification; in Track 2 it was generally secondary to model diversity and data coverage.

The official analysis also found that larger models tended to score better: Spearman’s $$\mathrm{BA}=\tfrac12\left(\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}+\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}\right)$$1 between peak VRAM usage and balanced accuracy was 0.616 with $$\mathrm{BA}=\tfrac12\left(\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}+\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}\right)$$2 (Aubreville et al., 5 Jun 2026). At the same time, multiple papers emphasize low-overhead deployment pathways. Percannella et al. prune auxiliary decoders at inference (Percannella et al., 28 Aug 2025), and “Mix, Align, Distil” applies all three added components only during training, leaving inference equivalent to a single DenseNet-121+CBAM forward pass (Atey et al., 28 Aug 2025). A plausible implication is that Track 2 performance gains often came from training-time regularization rather than from permanently more complex inference graphs.

6. Limitations, misconceptions, and future directions

A central limitation documented by the official challenge analysis is non-uniform generalization across biological contexts. Balanced accuracy varied significantly across tumor types, and false positive or false negative behavior remained domain dependent even in top systems (Aubreville et al., 5 Jun 2026). The hardest domains included feline gastrointestinal lymphoma and human glioblastoma, which the challenge paper describes as rare, highly pleomorphic tumors (Aubreville et al., 5 Jun 2026). This supports a cautious reading of aggregate balanced accuracy: strong pooled performance does not imply uniform reliability across all morphologies.

Another limitation concerns annotation difficulty. Kotte et al. report that 1,639 patches, or 13.7% of their Track 2 dataset, were “hard” instances with expert disagreement, including 1,064 AMF-H cases (Kotte et al., 1 Sep 2025). Their use of a dedicated hardness head indicates that a subset of the task difficulty is epistemic rather than purely representational. Similarly, the official challenge paper’s Cohen’s $$\mathrm{BA}=\tfrac12\left(\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}+\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}\right)$$3 confirms substantial but not perfect agreement among primary raters (Aubreville et al., 5 Jun 2026).

Several papers identify omitted or underexplored directions. Percannella et al. note that all loss weights were fixed to 1 and suggest that uncertainty weighting or grid search could yield further gains; they also propose self-supervised pretraining, meta-learning over domains, and additional cell-level auxiliary tasks (Percannella et al., 28 Aug 2025). Shen et al. explicitly state that no adversarial or feature-alignment loss was used, and point to targeted domain-adaptation as future work (Shen et al., 28 Aug 2025). Ochi and Bae note that LoRA rank was fixed and that FDA can introduce style artifacts (Ochi et al., 29 Aug 2025). Meng et al. highlight the inference overhead of stain-normalization TTA, which multiplies inference by a factor of 8 (Meng et al., 1 Sep 2025).

The challenge paper itself recommends additional AMF examples from low-prevalence domains, domain-balanced sampling strategies or domain-aware adapters such as conditional LoRA, and uncertainty quantification modules for deployment (Aubreville et al., 5 Jun 2026). Taken together, these recommendations indicate that MIDOG 2025 Track 2 established a strong benchmark for atypical mitosis classification under domain shift, but also that robust performance “in the wild” remains contingent on coverage of rare domains, careful class-imbalance handling, and methods that can separate true morphological signal from institution- or scanner-specific nuisance variation (Aubreville et al., 5 Jun 2026).