MIDOG++ Mitosis Detection Dataset

• MIDOG++ is a comprehensive pan-cancer dataset for mitosis detection, integrating data from seven diverse domains including varied tumor types and imaging protocols.
• It contains 11,937 annotated mitotic figures from 503 cases, facilitating domain-generalized studies in histopathology and robust model training.
• The dataset underpins protocols such as leave-one-domain-out validation and supports both detection and atypical classification with metrics like F1-score and balanced accuracy.

Searching arXiv for the specified MIDOG++-related papers to ground the article in the cited literature. MIDOG++ is a pan-cancer, multi-domain dataset for mitotic figure analysis in histopathology, introduced as an extension of the MIDOG 2021/2022 releases to support robust domain-generalized mitosis detection under variation in tumor type, laboratory workflow, scanner, and species. In the challenge literature, it serves primarily as the public training resource for mitosis detection, while later MIDOG 2025 resources derived from MIDOG++ also support atypical-versus-normal mitosis classification. Reports describing the dataset most commonly summarize it as containing 11,937 annotated mitotic figures across 503 histological specimens or individual tumor cases distributed over seven domains (Percannella et al., 28 Aug 2025).

1. Identity and intended role

MIDOG++ is explicitly described as a comprehensive, multi-domain dataset designed to advance robust, domain-generalized mitosis detection in histopathology. Multiple challenge reports identify it as the successor to MIDOG 2021 and MIDOG 2022, with the “++” denoting an expansion beyond the earlier emphasis on scanner variability toward broader variation in tumor morphology, laboratories, scanners, and species. In that framing, generalization rather than single-domain optimization is the central objective (Shen et al., 28 Aug 2025).

The dataset is used in the MIDOG 2025 ecosystem as the primary public detection resource, while the 2025 challenge itself adds a more realistic evaluation regime rather than a new public detection training set. A plausible implication is that MIDOG++ should be understood less as a static benchmark narrowly tied to one competition year and more as the common training substrate for successive domain-generalization studies in mitosis analysis (Aubreville et al., 5 Jun 2026).

2. Composition and domain structure

The core dataset is reported as comprising 503 histological specimens or individual tumor cases and 11,937 labeled or annotated mitotic figures, organized into 7 distinct domains. Each domain corresponds to a specific tumor type, acquired with a specific scanner and from a specific species, reflecting domain shifts due to biology, lab processing, and digitization. Several reports further describe MIDOG++ as extending the earlier MIDOG releases by broadening domain diversity in terms of tumor types, laboratories, whole-slide scanners, and species, making it the first dataset to incorporate such a wide domain shift across these factors (Percannella et al., 28 Aug 2025).

The tumor types listed for MIDOG++ are the following:

Tumor type
breast carcinoma
lung carcinoma
lymphosarcoma
neuroendocrine tumor
cutaneous mast cell tumor
cutaneous melanoma
(sub)cutaneous soft tissue sarcoma

MIDOG++ is described as consisting of region of interest images from histological specimens rather than full-slide dumps. At the same time, several task-specific papers do not restate scanner model names, magnification levels, microns-per-pixel, stain specifics, or file formats for the core dataset. This uneven reporting is a recurrent feature of the literature: the dataset is consistently treated as multi-laboratory and multi-scanner, but low-level acquisition metadata are often deferred to the original Scientific Data publication and official repository (Percannella et al., 28 Aug 2025).

3. Annotation regime and label semantics

The native task of MIDOG++ is mitotic figure detection. In at least one downstream study, the dataset is explicitly described as providing point annotations for mitotic figures, with the “center of each mitosis” taken from the dataset metadata. That study converts the point annotations to segmentation masks by passing each mitosis center to NuClick, thereby reframing detection as segmentation for training; this conversion is a modeling choice rather than the dataset’s native representation (Percannella et al., 28 Aug 2025).

The literature does not restate the annotation geometry uniformly. One challenge report notes that the Scientific Data paper provides “labels for 11,937 mitotic figures” but does not restate the exact label geometry in its own summary; another object-detection paper trains YOLO models and therefore treats the task operationally as bounding-box detection, but likewise does not detail the official annotation schema. This suggests that MIDOG++ is often consumed through task-specific preprocessing layers rather than through a single universally reiterated label interface (Shen et al., 28 Aug 2025).

The challenge literature also connects MIDOG++ to PHH3-assisted expert annotation. The MIDOG 2025 paper states that MIDOG++ uses PHH3-assisted annotation and majority voting, and the 2025 test annotations followed an adapted two-phase workflow in which H&E and PHH3 whole-slide images were manually registered in the EXACT tool, candidate cells were assigned to five categories, and disagreements in blinded review were resolved by a third pathologist. For the 2025 detection labeling, Cohen’s $\kappa \approx 0.48$ is reported, indicating moderate agreement in mitotic figure recognition (Aubreville et al., 5 Jun 2026).

A separate derivative annotation layer was released for atypical mitotic figure classification. In that setting, image patches of size $128 \times 128$ px centered on confirmed mitotic figures from MIDOG++ hotspots were labeled as normal versus atypical by two pathologists with third-expert tie-break resolution. For this atypical-versus-normal task, Cohen’s $\kappa \approx 0.68$ is reported (Aubreville et al., 5 Jun 2026).

4. Experimental protocols and evaluation criteria

A standard detection protocol built on MIDOG++ is leave-one-domain-out (LODO). In each run, 6 domains are used for training and 1 held-out domain is used for validation, explicitly simulating domain shift with the target domain unseen at training time. In that protocol, the source is the pooled training domains, the target is the held-out domain, and model selection is based on validation-domain loss. The reported validation metric is the F1-score on the held-out domain (Percannella et al., 28 Aug 2025).

For classification, a related but distinct protocol is leave-one-cancer-type-out cross-validation (LOOCV). In the EfficientViT study, for each held-out cancer type the model is trained on the other six types, and within each LOOCV setting 5-fold cross-validation is applied on the training pool, yielding 35 models across seven held-out types. Final probabilities are obtained by averaging predictions from the 5 models associated with the held-out type, and Otsu thresholding is used to select decision thresholds automatically per cancer type for threshold-dependent metrics (Qi et al., 28 Aug 2025).

Across these usages, the principal metrics are F1-score for detection and balanced accuracy for atypical-versus-normal classification. The formulas explicitly reported in the literature are

$$\mathrm{Precision} = \frac{TP}{TP + FP}, \qquad \mathrm{Recall} = \frac{TP}{TP + FN},$$

$$\mathrm{F1} = \frac{2 \cdot \mathrm{Precision} \cdot \mathrm{Recall}}{\mathrm{Precision} + \mathrm{Recall}},$$

and

$$\mathrm{BA} = \frac{1}{2}\left(\frac{TP}{TP+FN} + \frac{TN}{TN+FP}\right).$$

An important nuance is that not all MIDOG++-based papers restate the same detection-matching rule. The Mamba-based VM-UNet report focuses on F1 and does not specify the spatial matching threshold or additional detection metrics such as IoU, AP, or mAP. By contrast, the official MIDOG 2025 evaluation uses Hungarian matching with a maximum center-to-center distance of $7.5\,\mu m$, reports micro-averaged F1 as the main metric, and also includes FROC-AUC and Average Precision as secondary metrics (Percannella et al., 28 Aug 2025, Aubreville et al., 5 Jun 2026).

5. Preprocessing strategies and downstream operationalizations

A substantial part of the MIDOG++ literature concerns how the dataset is operationalized for training. In the VM-UNet study, region of interest images are tiled into $512 \times 512$ px patches with 80% overlap for both training and inference, and batches are balanced to contain equal numbers of positive tiles, defined as tiles with at least one mitosis, and negative tiles with no mitoses. Detection is reframed as segmentation by generating masks from mitosis centers, and inference outputs are converted back to detections by morphological dilation, connected-components analysis, and reporting the centers of minimum-area bounding boxes enclosing each component (Percannella et al., 28 Aug 2025).

That same study uses structure-preserving stain perturbation via Vahadane stain separation. With stain matrix $S$ and concentration matrix $C$ estimated for each image $128 \times 128$0, random scaling and shifting coefficients $128 \times 128$1 and $128 \times 128$2 are applied and the perturbed image is reconstructed as

$128 \times 128$3

The paper does not specify the ranges for $128 \times 128$4 and $128 \times 128$5, but the augmentation is explicitly intended to improve robustness against domain shift (Percannella et al., 28 Aug 2025).

For atypical mitosis classification, the EfficientViT study standardizes inputs to $128 \times 128$6 and applies H&E stain-based augmentation following Tellez et al. (2018): RGB patches are converted to H&E stain space via color deconvolution, each channel is perturbed independently by random scaling and shifting, and the result is projected back to RGB. Weighted sampling, focal loss, and AdamW are used to handle the strong class imbalance, with atypical mitoses comprising 15.6% overall in the unified MIDOG++ + AMi-Br training corpus used in that paper (Qi et al., 28 Aug 2025).

Object-detection studies use a different operationalization. The ensemble YOLO report trains on $128 \times 128$7 input tiles extracted from ROIs and augments them with hue, saturation, and brightness shifts, Gaussian blur, sharpening, Gaussian noise, Mosaic, and CutMix. In that work, YOLOv5 and YOLOv8 are treated as complementary one-stage detectors whose precision–recall trade-offs can be combined through ensembling (Kelam et al., 3 Sep 2025).

6. MIDOG++ in the MIDOG 2025 challenge ecosystem

MIDOG++ is the main public training resource behind the detection track of MIDOG 2025, but the 2025 evaluation itself goes beyond the original hotspot-centric paradigm. The challenge test set comprises 122 cases or whole-slide images across 12 tumor domains, with three non-overlapping ROIs of $128 \times 128$8 per case—hotspot, random, and challenging—yielding 365 ROIs in total. The random ROIs are intended as the primary proxy for whole-slide deployment, whereas the challenging ROIs are enriched in imposters and artifacts such as apoptotic or necrotic cells, hyperchromatic nuclei, inflammatory cells, surgical-margin ink, out-of-focus regions, and tissue folds (Aubreville et al., 5 Jun 2026).

The MIDOG 2025 classification track is built on a derivative MIDOG++ resource rather than on the original detection labels alone. The organizers released a public atypical mitotic figure training dataset comprising annotations for all 11,939 MIDOG++ mitotic figures, labeled as normal versus atypical, via Zenodo DOI 10.5281/zenodo.16044804. Other resources used in challenge submissions include stMIDOG++, LUNG-MITO, OMG-Octo, and OMG-Octo Atypical, but those are complementary rather than constitutive parts of the original MIDOG++ detection dataset (Shen et al., 28 Aug 2025, Aubreville et al., 5 Jun 2026).

This challenge framing materially changes the interpretation of MIDOG++. Earlier MIDOG editions were centered on scanner-domain generalization and hotspot regions, whereas the 2025 setting tests whether models trained on MIDOG++ remain stable across broader biological and contextual variation. A plausible implication is that MIDOG++ should be read not only as a dataset but also as the anchor point for an evolving benchmark family whose later evaluations increasingly emphasize clinically realistic tissue context (Aubreville et al., 5 Jun 2026).

7. Reported performance patterns, impact, and documentation limits

Published results on MIDOG++ consistently present it as a difficult domain-generalization benchmark. In leave-one-domain-out validation, the Mamba-based VM-UNet study reports mean $128 \times 128$9 standard deviation F1-scores of $\kappa \approx 0.68$0 for U-Net, $\kappa \approx 0.68$1 for VM-UNet, and $\kappa \approx 0.68$2 for VM-UNet with stain augmentation; on the MIDOG 2025 Track 1 preliminary test set, the same study reports an F1-score of $\kappa \approx 0.68$3 and explicitly notes “large room for improvement” (Percannella et al., 28 Aug 2025).

For atypical-versus-normal classification, the EfficientViT study uses a unified MIDOG++ + AMi-Br dataset of 13,938 annotated nuclei or patches and reports an overall preliminary-evaluation ROC AUC of $\kappa \approx 0.68$4, raw accuracy of $\kappa \approx 0.68$5, sensitivity of $\kappa \approx 0.68$6, specificity of $\kappa \approx 0.68$7, and balanced accuracy of $\kappa \approx 0.68$8. Its leave-one-cancer-type-out results vary materially by held-out type, with balanced accuracy ranging from $\kappa \approx 0.68$9 for canine lymphoma to $$\mathrm{Precision} = \frac{TP}{TP + FP}, \qquad \mathrm{Recall} = \frac{TP}{TP + FN},$$0 for canine cutaneous mast cell tumor, underscoring domain dependence even within a seven-type setting (Qi et al., 28 Aug 2025).

Internal validation in the ensemble YOLO study shows a related pattern: YOLOv5-l achieves Precision $$\mathrm{Precision} = \frac{TP}{TP + FP}, \qquad \mathrm{Recall} = \frac{TP}{TP + FN},$$1, Recall $$\mathrm{Precision} = \frac{TP}{TP + FP}, \qquad \mathrm{Recall} = \frac{TP}{TP + FN},$$2, and F1 $$\mathrm{Precision} = \frac{TP}{TP + FP}, \qquad \mathrm{Recall} = \frac{TP}{TP + FN},$$3; YOLOv8-m achieves Precision $$\mathrm{Precision} = \frac{TP}{TP + FP}, \qquad \mathrm{Recall} = \frac{TP}{TP + FN},$$4, Recall $$\mathrm{Precision} = \frac{TP}{TP + FP}, \qquad \mathrm{Recall} = \frac{TP}{TP + FN},$$5, and F1 $$\mathrm{Precision} = \frac{TP}{TP + FP}, \qquad \mathrm{Recall} = \frac{TP}{TP + FN},$$6; and the YOLOv5 + YOLOv8 ensemble reaches Precision $$\mathrm{Precision} = \frac{TP}{TP + FP}, \qquad \mathrm{Recall} = \frac{TP}{TP + FN},$$7, Recall $$\mathrm{Precision} = \frac{TP}{TP + FP}, \qquad \mathrm{Recall} = \frac{TP}{TP + FN},$$8, and F1 $$\mathrm{Precision} = \frac{TP}{TP + FP}, \qquad \mathrm{Recall} = \frac{TP}{TP + FN},$$9. The interpretation given there is that anchor-based and anchor-free detectors express complementary precision–recall behavior in sparse, heterogeneous mitosis settings (Kelam et al., 3 Sep 2025).

The official MIDOG 2025 analysis extends these observations to a more realistic test regime. Across participants, the best detection F1 reaches $$\mathrm{F1} = \frac{2 \cdot \mathrm{Precision} \cdot \mathrm{Recall}}{\mathrm{Precision} + \mathrm{Recall}},$$0 and the best atypical classification balanced accuracy reaches $$\mathrm{F1} = \frac{2 \cdot \mathrm{Precision} \cdot \mathrm{Recall}}{\mathrm{Precision} + \mathrm{Recall}},$$1. Mean detection performance falls from $$\mathrm{F1} = \frac{2 \cdot \mathrm{Precision} \cdot \mathrm{Recall}}{\mathrm{Precision} + \mathrm{Recall}},$$2 F1 in hotspots to $$\mathrm{F1} = \frac{2 \cdot \mathrm{Precision} \cdot \mathrm{Recall}}{\mathrm{Precision} + \mathrm{Recall}},$$3 in random areas and $$\mathrm{F1} = \frac{2 \cdot \mathrm{Precision} \cdot \mathrm{Recall}}{\mathrm{Precision} + \mathrm{Recall}},$$4 in challenging areas, with the approximately $$\mathrm{F1} = \frac{2 \cdot \mathrm{Precision} \cdot \mathrm{Recall}}{\mathrm{Precision} + \mathrm{Recall}},$$5 precision gap between hotspots and challenging ROIs corresponding to a 208% increase in false positive rate in challenging regions. The same paper reports mean gains of 1.5 percentage points in F1 from ensembling for detection and 1.3 percentage points in balanced accuracy for atypical classification, whereas test-time augmentation shows no relevant improvement (Aubreville et al., 5 Jun 2026).

Several documentation limits recur across the literature. Task-specific papers often omit scanner brands or models, magnification, microns-per-pixel, original patch-extraction protocol, file formats, focal-loss hyperparameters, or licensing details, and they do not consistently restate annotation geometry or inter-rater agreement for the original dataset. There is also a bibliographic inconsistency: multiple reports cite Aubreville et al., Scientific Data, 2023, DOI 10.1038/s41597-023-02327-4, whereas one YOLO report cites a 2024 Scientific Data article with DOI 10.1038/s41597-024-03548-x; the supplied reports do not reconcile this discrepancy (Shen et al., 28 Aug 2025, Kelam et al., 3 Sep 2025). This suggests that researchers using MIDOG++ should verify dataset metadata, access conditions, and exact release provenance against the official Scientific Data landing page and challenge repositories before implementation-level reuse.