MitoDetect++: Dual-Stage Pathology Pipeline

• The paper presents a dual-stage pipeline combining an attention-augmented U-Net-like detector and a Virchow2 transformer for robust mitosis detection and atypical classification.
• It employs strong augmentations, focal loss, group-aware stratified 5-fold cross-validation, and ensembling to address class imbalance and domain shift.
• The reported validation accuracy of 0.892 for atypical-versus-normal classification underlines its potential for automated mitotic counting in clinical settings.

MitoDetect++ is a domain-robust, dual-stage deep learning pipeline for computational pathology that was designed for the MIDOG 2025 challenge and targets two linked tasks: mitosis detection and atypical mitosis classification. In its published form, the system combines an attention-augmented U-Net-like detector with an EfficientNetV2-L encoder for Track 1 and a Virchow2 vision transformer fine-tuned with Low-Rank Adaptation for Track 2, while using strong augmentations, focal loss, group-aware stratified 5-fold cross-validation, ensembling, and test-time augmentation to mitigate class imbalance and domain shift. The reported headline result is an overall balanced accuracy of 0.892 for atypical-versus-normal classification across validation domains, with the broader design explicitly framed as suitable for automated mitotic counting and atypical mitosis quantification in clinical pipelines (Nasir et al., 28 Aug 2025).

1. Problem setting and challenge definition

MitoDetect++ addresses a problem space in which mitotic figures are both clinically important and technically difficult to model. Mitotic count is a core element of tumor grading and prognosis, while atypical mitoses such as tripolar, multipolar, asymmetrical, and lag-type forms are associated with genomic instability and more aggressive disease. The computational burden arises from extreme class imbalance, small target size, morphological variability, and domain shift caused by differences in labs, scanners, staining protocols, and tissue types (Nasir et al., 28 Aug 2025).

The MIDOG 2025 challenge formalized this setting into two tracks. Track 1 is mitosis detection, formulated here as binary segmentation of circular regions around mitosis centroids with diameter 21 px. Track 2 is atypical mitosis classification, formulated as binary classification of patches centered on annotated mitotic figures. For Track 2, the paper reports domain-wise sensitivity, specificity, accuracy, balanced accuracy, and AUROC, with balanced accuracy defined as

$\text{BA} = \frac{1}{2}(\text{TPR} + \text{TNR}).$

This formulation places MitoDetect++ in continuity with earlier MIDOG-era work that focused on domain generalization under scanner shift and related distributional changes (Razavi et al., 2021).

A common source of confusion is that the strongest numeric result reported for MitoDetect++ concerns Track 2 rather than Track 1. The paper gives detailed validation metrics for atypical-versus-normal classification, whereas for detection it primarily describes the architecture and training configuration underlying the challenge submission rather than a full detection metric table (Nasir et al., 28 Aug 2025).

2. Unified pipeline architecture

MitoDetect++ is organized as a two-stage unified pipeline. In the first stage, mitosis detection takes as input whole-slide images or large patches, extracts 512×512 patches at 0.25 µm/pixel, converts mitosis centroids into binary segmentation masks by dilation to circular blobs, and outputs a probability map of mitotic regions. In the second stage, atypical mitosis classification takes patches centered on previously annotated mitoses, processes them with a Virchow2 transformer backbone plus a binary head, and outputs the probability that a candidate is atypical rather than normal (Nasir et al., 28 Aug 2025).

This decomposition is methodologically aligned with a recurring pattern in the mitosis literature: a high-recall proposal stage followed by a more selective refinement stage. Earlier MIDOG systems implemented this logic with a fused detector and deep ensemble classifier (Liang et al., 2021), while "Mitosis Detection, Fast and Slow: Robust and Efficient Detection of Mitotic Figures" used candidate segmentation followed by EfficientNet-B7-based candidate refinement (Jahanifar et al., 2022). MitoDetect++ preserves the same systems-level intuition, but instantiates it as segmentation for Track 1 and transformer-based binary classification for Track 2.

The integration concept is direct. Track 1 proposes candidate mitoses in WSIs, and Track 2 classifies each candidate as atypical or normal. This suggests a slide-level workflow in which candidate generation, subtype assignment, and downstream quantification can be chained without requiring a single monolithic model (Nasir et al., 28 Aug 2025).

3. Detection component

The detection stage uses a U-Net-like encoder–decoder with an EfficientNetV2-L encoder pretrained on ImageNet. Decoder blocks use skip connections and include spatial and channel attention modules, and the final decoder layer outputs a single-channel map representing the probability of each pixel belonging to a mitosis region. Labels are produced by dilating mitosis centroids to disks with diameter 21 px, so the model is trained as a binary segmentation network rather than as a point detector or bounding-box detector (Nasir et al., 28 Aug 2025).

Training uses the MIDOG++ dataset with 10% held out as a test set, stratified by tissue type, and the remaining 90% processed with 5-fold cross-validation, also stratified by tissue type. CMC dataset patches are sampled during training to diversify domains. Optimization uses AdamW with initial learning rate $4 \times 10^{-4}$ and weight decay $0.01$. The loss is a combination of Jaccard, Dice, and Focal losses, and sampling is controlled by a RandomSampler that ensures at least 40% of patches per batch contain mitotic figures. Early stopping is triggered if validation loss does not decrease for 20 epochs. At inference, the top 3 checkpoints from internal 5-fold cross-validation are averaged (Nasir et al., 28 Aug 2025).

The detector therefore differs from earlier object-detection formulations such as Cascade R-CNN (Razavi et al., 2021), RF-DETR (Giedziun et al., 29 Aug 2025), and task-specific YOLO variants (Topuz et al., 1 Sep 2025), but it targets the same domain-robust small-object regime. A plausible implication is that segmentation was chosen here to encode local spatial context around centroid annotations more explicitly than anchor-based or query-based formulations.

4. Atypical mitosis classification component

The classification stage is built on Virchow2, a large vision transformer tailored for histopathology and trained in a self-supervised manner on massive pathology datasets. In MitoDetect++, the classification head is replaced with a single output neuron for binary classification, with images resized to 224×224. Rather than full fine-tuning, the model uses Low-Rank Adaptation, parameterizing weight updates as

$\Delta W = A B^T,$

with rank $r = 8$, scaling factor $\alpha = 16$, and LoRA dropout $0.3$. LoRA modules are inserted into qkv projections, the attention output projection, and both fully connected layers $fc1$ and $fc2$ in the transformer MLP blocks (Nasir et al., 28 Aug 2025).

Training data for Track 2 are drawn from the MIDOG 2025 Atypical Set, the AMi-Br Breast Cancer dataset, the Zhuoyan Shen et al. dataset, and the Mitosis subtyping dataset. Splitting uses StratifiedGroupKFold, with slide identifiers as groups and atypical-versus-normal labels balanced across folds. This prevents slide-level leakage while stabilizing class proportions across validation folds (Nasir et al., 28 Aug 2025).

Optimization uses Adam with learning rate $5 \times 10^{-5}$, weight decay $4 \times 10^{-4}$0, and focal loss with $4 \times 10^{-4}$1 and $4 \times 10^{-4}$2. Class imbalance is further addressed with WeightedRandomSampler inversely proportional to class frequencies and early stopping after 10 epochs without validation improvement. Inference merges LoRA weights into the base Virchow2 network, then applies test-time augmentation with scaling $4 \times 10^{-4}$3, flips, 90° rotations, and brightness adjustments, followed by averaging across TTA variants and folds (Nasir et al., 28 Aug 2025).

This design is consistent with prior evidence that automatic subtyping of mitotic figures is feasible but sensitive to label noise and inter-rater disagreement. A hierarchical anchor-free FCOS-based system previously reported ROC AUC 0.833 for atypical-versus-normal discrimination and mean class-averaged ROC-AUC 0.977 for phase discrimination, establishing atypical subtyping as a viable extension of mitosis detection rather than a fundamentally separate task (Aubreville et al., 2022).

5. Domain-robustness mechanisms and reported performance

MitoDetect++ uses several explicit domain-robustness mechanisms. Strong data augmentation targets scanner noise, tissue orientation, and staining variability; focal loss addresses class imbalance in both segmentation and classification; group-aware stratified cross-validation enforces slide-level independence; test-time augmentation reduces sensitivity to appearance shifts; and ensembling averages predictions from multiple checkpoints or folds to reduce variance (Nasir et al., 28 Aug 2025).

The reported validation results for Track 2 are domain-wise rather than aggregate-only, which is important because robustness is the central claim. The overall values are AUC 0.964, accuracy 0.853, sensitivity 0.958, specificity 0.827, and balanced accuracy 0.892. Domain 3 is near-perfect, while Domain 0 is substantially harder, especially in sensitivity. This distribution suggests that robustness is strong but not uniform across all domains (Nasir et al., 28 Aug 2025).

Domain	AUC / Acc	Sens / Spec / BA
0	0.820 / 0.750	0.500 / 0.781 / 0.641
1	0.954 / 0.851	0.966 / 0.826 / 0.896
2	0.985 / 0.864	1.000 / 0.809 / 0.905
3	1.000 / 0.921	1.000 / 0.917 / 0.958
Overall	0.964 / 0.853	0.958 / 0.827 / 0.892

The lower sensitivity in Domain 0 indicates that rare atypical instances in harder domains may still be missed. The paper explicitly links this to possible domain shift or fewer atypical examples and suggests that domain-specific augmentation or adaptation may still be needed (Nasir et al., 28 Aug 2025).

For contextualization, other MIDOG-style domain-robust detectors have reported F1 values in a similar benchmark space, but for detection rather than atypical subtyping: 0.7492 for a Cascade R-CNN method (Razavi et al., 2021), 0.7550 for a fused detector plus deep ensemble classifier (Liang et al., 2021), 0.766 for SDF-YOLO (Topuz et al., 1 Sep 2025), and 0.789 for RF-DETR (Giedziun et al., 29 Aug 2025). These figures are not direct substitutes for the Track 2 balanced accuracy of MitoDetect++, but they indicate the broader performance range of contemporary domain-generalized mitosis detection systems.

6. Relation to prior systems, limitations, and future directions

MitoDetect++ belongs to a broader lineage of mitosis systems that alternately emphasize segmentation, object detection, or two-stage verification. Segmentation-oriented formulations include UV-Net, which preserved high-resolution features and reached F1 0.6721 on MIDOG 2021 (Mirjahanmardi et al., 2021), and the EUNet-based fast/slow framework that won MIDOG21 and MIDOG22 (Jahanifar et al., 2022). Detection-oriented formulations include Cascade R-CNN (Razavi et al., 2021), FoCasNet’s hybrid-anchor two-stage detector with F1 0.888 on ICPR 2012 (Wang et al., 2023), MitDet’s diversity-guided point-annotation framework (Wang et al., 2023), and several 2025 one-stage or transformer detectors (Topuz et al., 1 Sep 2025). MitoDetect++ is distinctive in unifying detection and atypical subtyping within one challenge-specific pipeline rather than treating classification as an auxiliary study (Nasir et al., 28 Aug 2025).

Its clinical applicability follows directly from the workflow. Track 1 can run on WSIs to identify candidate mitotic figures, Track 2 can classify each candidate as normal or atypical, and the outputs can be visualized in digital pathology systems as overlays or lists of atypical mitoses. The use of LoRA makes the Virchow2 stage resource-efficient, while patch-based detection remains compatible with parallel WSI tiling. This suggests scalability across tasks and deployment settings, although the ensemble size and transformer backbone still imply nontrivial runtime costs (Nasir et al., 28 Aug 2025).

The principal limitations are also explicit. Domain 0 sensitivity is lower than in the other domains. Rare atypical subtypes and label noise remain intrinsic constraints. Dataset composition, although broader than a single-center study, may still bias the model toward certain tissue types or centers. Another important interpretive point is that the paper does not present detailed Track 1 numeric detection metrics in the same way it reports Track 2 validation metrics; the balanced accuracy 0.892 should therefore not be read as a detection score (Nasir et al., 28 Aug 2025).

The future directions suggested are domain adversarial training, few-shot learning, better domain adaptation for difficult domains, extension to tasks such as grading, necrosis detection, and nuclear atypia, and integration with larger vision-LLMs for explainable AI or report generation. A plausible implication is that MitoDetect++ functions less as a terminal architecture than as a modular template: segmentation for candidate generation, efficient transformer adaptation for subtype discrimination, and domain-robust training/inference schemes that can be transferred to adjacent histopathology tasks (Nasir et al., 28 Aug 2025).