MIDOG 2025 Task 2: Atypical Mitosis Classification

• MIDOG 2025 Task 2 is a binary classification task that distinguishes normal mitotic figures from atypical ones, addressing domain shifts from varying tumor types, species, and staining protocols.
• The task leverages multi-domain data with diverse annotations and a significant class imbalance to improve domain-robust fine-grained morphology recognition.
• Methodologies range from modern convolutional networks to transformer-based models, using balanced accuracy and advanced augmentation techniques to overcome morphological ambiguity.

Searching arXiv for MIDOG 2025 Task 2 papers to ground the article in the latest challenge literature. MIDOG 2025 Task 2 is the atypical mitotic figure classification track of the Mitosis Domain Generalization challenge, formulated as a binary classification problem on image patches centered on confirmed mitotic figures. The task requires assigning each patch to either a normal mitotic figure (NMF) or an atypical mitotic figure (AMF) under substantial domain shift arising from differences in tumor type, species, laboratory, staining protocol, and scanner. Across the 2025 literature, Task 2 is consistently framed as a test of domain-robust fine-grained morphology recognition, with balanced accuracy as the principal metric because AMFs are a minority class and because cross-domain robustness, rather than raw accuracy, is the organizing challenge (Yamagishi et al., 26 Aug 2025).

1. Definition and challenge scope

MIDOG 2025 comprises two tracks: robust mitotic figure detection and atypical mitotic figure classification. Task 2 corresponds to the latter and is defined as follows: given an image patch centered on a mitotic figure, classify the mitosis as typical or atypical, equivalently normal or atypical, under strong variation in scanners, centers, tissues, tumor types, and species (Percannella et al., 28 Aug 2025). The input format is patch-level rather than whole-slide level, and the output is binary, typically encoded as class 0 for NMF and class 1 for AMF, or equivalently typical versus atypical (Yamagishi et al., 26 Aug 2025).

The clinical motivation stated in the challenge literature is that atypical mitoses reflect abnormal division patterns and are associated with tumor aggressiveness, genomic instability, and prognostic relevance. The computational difficulty is not the presence of a mitotic figure per se, since Track 2 patches are centered on confirmed mitoses, but the discrimination of subtle morphological irregularities from the range of appearances exhibited by normal mitotic phases (Percannella et al., 28 Aug 2025). This is further complicated by the fact that both normal and atypical mitoses display high intra-class variability, while their inter-class differences are often visually subtle (Yamagishi et al., 26 Aug 2025).

The domain generalization aspect is explicit. The challenge evaluates performance on unseen domains, and the domain concept itself is multi-factorial: scanner, laboratory or origin, species, and tumor type are recurrently treated as components of domain identity (Atey et al., 28 Aug 2025). A plausible implication is that Task 2 is best understood not merely as binary classification, but as binary classification under structured distribution shift.

2. Data resources, annotations, and label structure

The official MIDOG 2025 atypical training data are described in several closely related but not identical ways across submissions. One formulation reports 11,939 mitotic figures from the MIDOG 2025 Atypical Training Set, with 10,191 normal mitotic figures and 1,748 atypical mitotic figures, derived from 454 labeled images and nine distinct domains, with a test set covering 120 cases and 12 distinct tumor types including both human and veterinary pathology (Yamagishi et al., 26 Aug 2025). Another describes the MIDOG 2025 Atypical Training Set as encompassing 11,939 mitotic figures from all seven domains of MIDOG++ (Percannella et al., 28 Aug 2025). A third, post-challenge synthesis states that for MIDOG 2025 all 11,939 mitotic figures in MIDOG++ were re-annotated as AMF versus NMF and released as a public training dataset, while the final test set comprised hotspot-region mitotic figures from 122 whole-slide images across 12 tumor domains (Aubreville et al., 5 Jun 2026). The differing domain counts in the challenge papers reflect differences in how domain partitions were described in individual submissions rather than a single harmonized convention.

The available public ecosystem extends beyond the official training set. Frequently reused external resources include AMi-Br, OMG-Octo Atypical, AtNorM-Br, LUNG-MITO, and Jahanifar’s mitosis subtyping dataset (Ochi et al., 29 Aug 2025, Feki et al., 29 Aug 2025, Giedziun et al., 29 Aug 2025). These additions are used to enlarge morphological coverage and domain diversity. For example, AMi-Br is repeatedly described as containing 3,720 mitotic figures, of which 832 are atypical and 2,888 normal (Percannella et al., 28 Aug 2025). OMG-Octo Atypical is explicitly described as a refinement of the original OMG-Octo database to incorporate atypical mitosis labels (Shen et al., 28 Aug 2025).

The annotation structure is also important. In the official training set, labels are image-level binary subtype labels per mitotic figure crop (Yamagishi et al., 26 Aug 2025). Some works further exploit multi-expert annotations. One study converts three pathologist annotations into soft labels

$\tilde{y}_i = \frac{1}{3}\sum_{k=1}^3 y_i^{(k)},$

so that expert disagreement is represented probabilistically rather than collapsed immediately to a hard class (Giedziun et al., 29 Aug 2025). Another paper formalizes a separate notion of instance difficulty by designating patches with full expert agreement as easy and those with disagreement as hard, producing a secondary hardness label in addition to phenotype (Kotte et al., 1 Sep 2025). This suggests that label uncertainty itself is treated as a meaningful signal in some Task 2 formulations.

3. Morphological and statistical difficulties

Three recurrent properties define the difficulty of MIDOG 2025 Task 2. First is class imbalance. In the official training set, AMFs are a minority; one paper characterizes them as approximately 20% of all mitotic figures, while also providing the explicit ratio

$$\text{AMF proportion} \approx \frac{1748}{10191 + 1748} \approx 0.146$$

and noting that “$\sim 20\%$” is a rounded characterization (Yamagishi et al., 26 Aug 2025). Another combined dataset after deduplication reports 1,771 AMFs and 10,168 NMFs, a class ratio of approximately $1:5.74$ (Giedziun et al., 29 Aug 2025). The challenge overview emphasizes that AMFs remain a minority in nearly all domains of the final test set (Aubreville et al., 5 Jun 2026).

Second is morphological ambiguity. Atypical mitoses are described as cells that do not follow a structured division pattern and may exhibit tripolar or multipolar spindles, asymmetrical chromatin distribution, lagging chromosomes, ring-shaped mitoses, or fragmented chromosome configurations (Percannella et al., 28 Aug 2025, Aubreville et al., 5 Jun 2026). Normal mitoses encompass the canonical phases of prophase, metaphase, anaphase, and telophase (Percannella et al., 28 Aug 2025). The challenge papers repeatedly note that the distinction between NMF and AMF can be subtle even for experts, and that inter-observer variability is substantial (Meng et al., 1 Sep 2025, Balezo et al., 28 Aug 2025).

Third is domain heterogeneity. Differences in scanner hardware, acquisition pipeline, stain appearance, tissue morphology, and species affect color distribution, texture, background microenvironment, and artefact patterns (Percannella et al., 28 Aug 2025). Post-challenge analysis found significant variation in Track 2 performance across tumor domains, with human astrocytoma among the easiest and feline GI lymphoma among the hardest, indicating persistent domain-specific blind spots even in top-performing models (Aubreville et al., 5 Jun 2026).

4. Predominant methodological families

Task 2 methods fall into a small number of clearly defined technical families.

The first family is modern convolutional classification. ConvNeXt is especially prominent. One submission used ConvNeXt V2 Base with a single-neuron output head, BCEWithLogitsLoss, 60% center cropping, five-fold stratified cross-validation, and fold averaging, achieving a cross-validation balanced accuracy of $0.8314 \pm 0.0261$ and development-phase balanced accuracy of 0.8831 (Yamagishi et al., 26 Aug 2025). Another large comparative study found ConvNeXt-large to be the best single model among ViT, ResNet, DenseNet, ConvNeXt, and EfficientNet variants, with cross-validation balanced accuracy $0.8794 \pm 0.0082$, and used a ConvNeXt ensemble to achieve preliminary balanced accuracy 0.86 (Xu et al., 29 Aug 2025). A further ConvNeXt Small–based study trained on AMi-Br, AtNorM-Br, MIDOG++, and OMG-Octo, combining weighted sampling with extensive histopathology-specific augmentation, and reported preliminary leaderboard balanced accuracy 0.8961 and ROC AUC 0.9561 (Feki et al., 29 Aug 2025).

The second family is pathology foundation models with parameter-efficient adaptation. Virchow, Virchow2, UNI, UNI2-h, H-optimus-0, and HIBOU all appear in the 2025 literature. LoRA is the dominant PEFT mechanism. Ramchandani et al. evaluated Virchow, Virchow2, and UNI with LoRA and found their best approach, Virchow with rank-8 LoRA and three-fold ensembling, reached balanced accuracy 88.37% on the preliminary test set (Ramchandani et al., 21 Sep 2025). A DINOv3-H+ model pretrained on natural images and adapted with LoRA on query and value projections, using only about 650k trainable parameters, achieved preliminary balanced accuracy 0.8871 (Balezo et al., 28 Aug 2025). Another foundation-model submission used H-optimus-0 with LoRA, soft labels, adaptive focal loss, hard negative mining, supervised contrastive regularization, and a domain-adversarial head, achieving mean leave-one-domain-out balanced accuracy $0.851 \pm 0.037$ across 10 domains (Giedziun et al., 29 Aug 2025).

The third family is prompt-based or other parameter-efficient adaptation of pathology transformers. Meng et al. found that visual prompt tuning on UNI2-h improved over a LoRA baseline, increasing balanced accuracy from 0.8305 to 0.8711, and that adding TTA with Vahadane and Macenko stain normalization further raised it to 0.8837 with ROC AUC 0.9513 (Meng et al., 1 Sep 2025).

The fourth family is multi-task or object-centric learning. Percannella et al. proposed a multi-task network with a shared Pre-Activated ResNet-50 backbone and three heads: image-level classification, binary mitosis segmentation, and pixel-level mitosis classification. Under leave-one-domain-out evaluation, the multi-task model improved balanced accuracy over a single-task baseline on validation, unseen test domains, and external AMi-Br, and achieved 0.856 on the preliminary MIDOG25 test set (Percannella et al., 28 Aug 2025). A separate teacher–student framework coupled UNet-based segmentation, contrastive learning, domain-adversarial training, and a multi-scale classifier head, reaching 0.8418 balanced accuracy on the preliminary test set for Track 2 (Choe et al., 3 Sep 2025).

The fifth family emphasizes domain generalization through explicit training-time regularization. “Mix, Align, Distil” combines MixStyle in early and mid DenseNet-121 stages, CBAM-based cross-domain alignment using weak metadata, and EMA-teacher distillation with temperature-scaled KL divergence. Its preliminary leaderboard results were balanced accuracy 0.8762, sensitivity 0.8873, specificity 0.8651, and ROC AUC 0.9499 (Atey et al., 28 Aug 2025).

5. Training strategies, imbalance handling, and evaluation protocol

Despite substantial architectural diversity, several training practices recur.

Balanced accuracy is the dominant primary metric. It is typically defined as

$BA = \frac{TPR + TNR}{2},$

with $TPR = \frac{TP}{TP+FN}$ and $TNR = \frac{TN}{TN+FP}$ (Yamagishi et al., 26 Aug 2025, Percannella et al., 28 Aug 2025). Challenge-wide reporting also includes ROC AUC, and many papers additionally report sensitivity and specificity (Yamagishi et al., 26 Aug 2025, Ramchandani et al., 21 Sep 2025, Meng et al., 1 Sep 2025).

Cross-validation protocols vary. Stratified five-fold cross-validation is common in pure classifier pipelines (Yamagishi et al., 26 Aug 2025, Feki et al., 29 Aug 2025, Kotte et al., 1 Sep 2025). Leave-one-domain-out is used when domain shift itself is the principal experimental variable (Percannella et al., 28 Aug 2025, Giedziun et al., 29 Aug 2025). Some submissions combine the official MIDOG training set with external datasets and perform four-fold cross-validation on the merged corpus (Xu et al., 29 Aug 2025, Balezo et al., 28 Aug 2025).

Loss design reflects class imbalance. BCEWithLogitsLoss is frequently used without explicit class weights in ConvNeXt pipelines (Yamagishi et al., 26 Aug 2025, Xu et al., 29 Aug 2025). Other methods explicitly address imbalance through focal loss (Balezo et al., 28 Aug 2025, Kotte et al., 1 Sep 2025), adaptive focal loss with dynamic positive weighting (Giedziun et al., 29 Aug 2025), or weighted binary cross-entropy based on inverse class frequency (Percannella et al., 28 Aug 2025). Sampling-based corrections are also widespread: WeightedRandomSampler is used both for balancing classes (Ramchandani et al., 21 Sep 2025, Feki et al., 29 Aug 2025) and for weighting datasets or domains (Ochi et al., 29 Aug 2025, Giedziun et al., 29 Aug 2025).

Augmentation is central to domain robustness. Geometric transforms such as flips, rotations, random crops, shift-scale-rotate, elastic transforms, grid distortion, and perspective warping are common (Yamagishi et al., 26 Aug 2025, Feki et al., 29 Aug 2025, Atey et al., 28 Aug 2025). Histopathology-specific color processing includes stain augmentation, ColorJitter, hue-saturation perturbations, CLAHE, defocus blur, Gaussian noise, ISO noise, and channel perturbations (Feki et al., 29 Aug 2025, Balezo et al., 28 Aug 2025). Several methods incorporate stain normalization at training or test time using Macenko or Vahadane (Meng et al., 1 Sep 2025). Fourier Domain Adaptation and fisheye transformations are also used in at least one pathology foundation model ensemble (Ochi et al., 29 Aug 2025).

One especially influential preprocessing choice is center cropping. In the ConvNeXt V2 submission, a 60% center crop reduced a $$\text{AMF proportion} \approx \frac{1748}{10191 + 1748} \approx 0.146$$0 patch to approximately $$\text{AMF proportion} \approx \frac{1748}{10191 + 1748} \approx 0.146$$1 before resizing back to 128. This improved training stability and raised balanced accuracy by $$\text{AMF proportion} \approx \frac{1748}{10191 + 1748} \approx 0.146$$2 relative to a no-cropping condition in converged folds, largely through a sensitivity gain of $$\text{AMF proportion} \approx \frac{1748}{10191 + 1748} \approx 0.146$$3 for AMFs (Yamagishi et al., 26 Aug 2025).

6. Performance patterns and post-challenge synthesis

The post-challenge organizer analysis provides the clearest global picture. MIDOG 2025 Track 2 received 21 submissions, and balanced accuracy on the final test set reached up to 0.908 (Aubreville et al., 5 Jun 2026). The winning method was a DINOv3-H model with LoRA adaptation trained on MIDOG 2025, AMi-Br, and OMG-Octo Atypical, obtaining balanced accuracy 0.908 without ensembling, while the strongest ROC AUC, 0.971, was reported by a ConvNeXt v2–based ensemble (Aubreville et al., 5 Jun 2026). The official baseline, an EfficientNetV2 classifier, achieved balanced accuracy 0.827 (Aubreville et al., 5 Jun 2026).

A notable result of the organizer ablation is that ensembling improves Track 2 performance on average by approximately 1.3 percentage points in balanced accuracy, whereas TTA shows no relevant improvement, with only about 0.3 percentage points mean gain (Aubreville et al., 5 Jun 2026). This is consistent with several individual papers in which ensemble averaging or majority voting is treated as a major robustness mechanism (Yamagishi et al., 26 Aug 2025, Percannella et al., 28 Aug 2025, Xu et al., 29 Aug 2025, Ochi et al., 29 Aug 2025).

Performance is not uniform across tumor types. Organizers report significant differences across the 12 tumor domains, with “blind spots” in rare or highly pleomorphic malignancies (Aubreville et al., 5 Jun 2026). Human astrocytoma shows the highest median balanced accuracy, while feline GI lymphoma shows the lowest, with reduced AMF recall (Aubreville et al., 5 Jun 2026). This suggests that cross-domain generalization in Task 2 remains limited by the representational coverage of atypical morphologies rather than by a single universal failure mode.

At the individual-method level, strong preliminary performances cluster in the high-0.87 to low-0.89 balanced-accuracy range for many competitive systems: 0.8837 for Virchow plus LoRA (Ramchandani et al., 21 Sep 2025), 0.8837 for VPT plus UNI2-h with stain-normalization TTA (Meng et al., 1 Sep 2025), 0.8871 for DINOv3-H+ with LoRA (Balezo et al., 28 Aug 2025), 0.8961 for ConvNeXt Small with histopathology-specific augmentation (Feki et al., 29 Aug 2025), and 0.9107 in one data-centric pan-cancer submission that aggregated five datasets and emphasized stain-aware augmentation and model selection around ConvNeXt-tiny (Shen et al., 28 Aug 2025). These figures indicate that the field converged quickly on a relatively narrow performance band, with differences often traceable to dataset breadth, augmentation policy, and adaptation mechanism rather than to a single dominant backbone class.

7. Limitations, controversies, and future directions

The challenge literature repeatedly identifies three unresolved issues.

The first is incomplete domain generalization. Even top-performing models exhibit marked domain-specific variability (Aubreville et al., 5 Jun 2026). Some methods explicitly acknowledge that their gains under domain shift are modest, for example the multi-task network that improves balanced accuracy by approximately 0.007 on unseen domains relative to a single-task baseline (Percannella et al., 28 Aug 2025). A plausible implication is that many current methods reduce but do not eliminate dependence on stain, scanner, or tissue-context cues.

The second is interpretability. Several papers note that these systems remain black-box classifiers with limited interpretability regarding the morphological features driving AMF versus NMF decisions (Yamagishi et al., 26 Aug 2025). This matters because AMF classification is clinically consequential and already subject to expert disagreement. Some future directions proposed in the literature include Grad-CAM, saliency methods, or concept-based analysis, but these are generally presented as prospective additions rather than evaluated components (Yamagishi et al., 26 Aug 2025, Giedziun et al., 29 Aug 2025).

The third is the status of atypical morphology itself as a label space. Most submissions treat Task 2 as binary subtype classification, but the morphological guide underlying the annotations distinguishes multiple abnormal patterns such as multipolarity, lagging chromosomes, and anaphase bridges (Aubreville et al., 5 Jun 2026). Several papers suggest, implicitly or explicitly, that finer-grained AMF subclassification could be a future extension (Shen et al., 28 Aug 2025, Aubreville et al., 5 Jun 2026). This suggests that the current binary formulation is both practical and reductive: practical because it provides a tractable benchmark, reductive because it compresses morphologically diverse abnormalities into a single positive class.

A further methodological tension concerns whether task-specific, lightweight CNNs or very large foundation models are preferable. Some submissions report that compact ConvNeXt variants or even ConvNeXt-tiny trained from scratch outperform more elaborate foundation-model pipelines when sufficient multi-domain data are assembled (Shen et al., 28 Aug 2025). Others show clear benefits from Virchow, UNI, or DINOv3 with LoRA or VPT (Ramchandani et al., 21 Sep 2025, Meng et al., 1 Sep 2025, Balezo et al., 28 Aug 2025). The challenge-level results indicate that both paradigms are competitive. This suggests that success in MIDOG 2025 Task 2 depends at least as much on data composition, augmentation, and calibration as on the choice between CNN and transformer backbones.

Overall, MIDOG 2025 Task 2 established atypical mitotic figure classification as a distinct, reproducible benchmark in computational pathology, moving beyond mitosis detection toward morphology-aware prognostic phenotyping. The challenge demonstrated that balanced accuracies near 0.91 are achievable on a multi-domain benchmark, yet also showed that rare tumor types, pleomorphic domains, and annotation ambiguity remain major constraints on generalization (Aubreville et al., 5 Jun 2026).