---
title: 'MIDOG 2025: Domain-Generalized Mitotic Detection'
url: https://www.emergentmind.com/topics/midog-2025-track-1
type: topic
---

# MIDOG 2025: Domain-Generalized Mitotic Detection

MIDOG 2025 Track 1 is the mitotic figure detection task of the MItosis DOmain Generalization challenge, designed to test whether computational pathology systems can localize mitotic figures robustly under domain shift rather than only within curated hotspots. In contrast to earlier MIDOG editions, the 2025 benchmark explicitly extends generalization pressure from scanner variation to broad biological and contextual heterogeneity, including multiple species, tumor types, scanners, and three region-of-interest contexts—hotspot, random, and challenging—thereby approximating whole-slide deployment conditions more closely than hotspot-only evaluation [2606.07368].

## 1. Scope and conceptual framing

MIDOG 2025 Track 1 addresses object detection of mitotic figures in H\&E-stained histopathology under pronounced domain shift. Across the solution papers, domain shift is consistently described in terms of scanner variability, stain variability, inter-site differences, tissue-context heterogeneity, species differences, and tumor-type morphology changes [2509.02593][2509.02597][2509.02599]. The challenge paper formalizes this expansion as a move toward detection “in the wild,” where algorithms must operate not only in highly proliferative tumor hotspots but also in random tissue areas and deliberately difficult regions containing artifacts and mitotic imposters [2606.07368].

This 2025 formulation differs materially from prior MIDOG editions. MIDOG 2021 primarily emphasized scanner-induced shift, whereas MIDOG 2022 broadened evaluation to multiple tumor types and species but remained hotspot-centered. MIDOG 2025 extends the generalization target further by combining twelve tumor types with three ROI contexts, making contextual robustness a first-class evaluation variable [2606.07368].

A central implication of this design is that Track 1 is not merely a small-object detection benchmark. It is also a robustness benchmark in which false positives from apoptotic or necrotic cells, inflammatory cells, hyperchromatic nuclei, focus artifacts, tissue folds, ink marks, and other hard negatives become decisive, especially outside hotspots [2606.07368]. Several teams therefore framed the task as a balance between aggressive candidate harvesting and equally aggressive false-positive control, often via two-stage pipelines or segmentation-based formulations [2509.02597][2509.02600][2509.03614].

## 2. Dataset, ROI construction, and annotation protocol

The final Track 1 test set comprised 122 whole-slide images, one per patient or case, across 12 tumor types and three ROI types per WSI, yielding 365 ROIs total: 122 hotspot ROIs, 121 random ROIs, and 122 challenging ROIs. One WSI lacked sufficient tissue for a non-overlapping random ROI [2606.07368]. The tumor domains were human melanoma, astrocytoma, bladder carcinoma, meningioma, colon carcinoma, glioblastoma, and lung adenocarcinoma; canine mammary carcinoma, cutaneous mast cell tumor, and hemangiosarcoma; and feline soft tissue sarcoma and gastrointestinal lymphoma [2606.07368].

The scanning platforms were Hamamatsu NanoZoomer S60 at 0.22 \(\mu\)m/px, Hamamatsu NanoZoomer S360 at 0.23 \(\mu\)m/px, 3DHISTECH Pannoramic Scan II at 0.25 \(\mu\)m/px, and 3DHISTECH P1000 at 0.12 \(\mu\)m/px for human glioblastoma and lung adenocarcinoma [2606.07368]. Each ROI covered 2 mm\(^2\).

| ROI type | Count | Selection principle |
|---|---:|---|
| Hotspot | 122 | Pathologist-selected for high cellular density and image/tissue quality |
| Random | 121 | Uniformly sampled from tissue area with minimum 80% tissue coverage |
| Challenging | 122 | Pathologist-curated for imposters and artifacts |

Annotation used a PHH3-assisted two-phase pipeline. A primary expert labeled five categories in registered H\&E–PHH3 views: PHH3-positive and clearly mitotic in H\&E, PHH3-positive and suggestive of mitosis, PHH3-negative but suggestive of mitosis, PHH3-positive but suggestive of an imposter, and PHH3-negative morphology resembling but distinct from mitosis. Classes \(1\)–\(3\) were counted as mitotic figure annotations, while classes \(4\)–\(5\) were retained as hard negatives [2606.07368]. A secondary expert then reviewed 30-\(\mu\)m H\&E patches centered on annotations, blinded to prior labels, and disagreements were resolved by a third expert [2606.07368].

This protocol is important because Track 1 does not reduce to binary foreground localization. The benchmark explicitly preserves hard negatives as a structured source of error pressure. Several participating methods adopted this signal directly by training with separate “hard negative” classes or by converting point annotations into segmentation masks that preserved non-mitotic lookalikes within the supervision regime [2509.02593][2509.03614]. Inter-observer agreement for Track 1 initial raters was Cohen’s \(\kappa = 0.48\), indicating moderate agreement and underscoring the intrinsic ambiguity of the task [2606.07368].

No new Track 1 training set was released for the final challenge. Participants trained on public data sources such as MIDOG++, MITOS\_WSI\_CMC, and MITOS\_WSI\_CCMCT, sometimes augmented with other public pathology datasets [2606.07368].

## 3. Evaluation protocol and metric interpretation

Track 1 evaluation used Hungarian assignment between predictions and mitotic figure ground truth with a maximum distance of \(7.5\,\mu\)m, approximately a nuclear diameter. Multiple detections of the same mitotic figure counted as false positives [2606.07368]. Evaluation was performed per ROI and reported separately for hotspot, random, and challenging contexts as well as in aggregate [2606.07368].

The primary metric was micro-averaged \(F_1\), computed as
\[
F_1 = \frac{2TP}{2TP + FN + FP}.
\]
Precision and recall were defined conventionally as
\[
\mathrm{Precision} = \frac{TP}{TP + FP}, \qquad
\mathrm{Recall} = \frac{TP}{TP + FN}.
\]
Secondary metrics were FROC-AUC and average precision. FROC-AUC was computed from sensitivity versus false-positives-per-image over \([0,8]\) FPPI, with 40 thresholds interpolated to 50 evaluation points, while AP was computed as mean precision across 101 uniformly spaced recall points using the torchmetrics implementation [2606.07368].

A frequent source of confusion in the MIDOG 2025 literature is the coexistence of multiple evaluation regimes. The official final leaderboard used the withheld 122-case, 365-ROI multi-context test set [2606.07368]. By contrast, many solution papers reported results on the preliminary test set, on internal MIDOG++ validation splits, or on fused-dataset train/validation/test splits. For example, a YOLOv12-based detector reported \(F_1 = 0.801\) on the preliminary test set [2509.02593], an RF-DETR submission reported \(F_1 = 0.789\) on the preliminary test set [2509.02599], an FCOS cascade reported \(F_1 = 0.8432\) on MIDOG++ validation [2509.02597], and an improved YOLO11x plus ConvNeXt framework reported \(F_1 = 0.882\) on a fused-dataset split rather than the official challenge test set [2509.02627]. These figures are therefore not directly interchangeable with the final leaderboard values.

The preliminary test set itself was hotspot-only and consisted of 20 cases across four tumor domains for Docker validation [2606.07368]. This design detail helps explain why methods that appeared strong during development could still degrade sharply on the final benchmark, where random and challenging ROIs were explicitly included.

## 4. Methodological landscape

The submitted methods spanned one-stage detection, two-stage detection-classification cascades, and segmentation-to-detection formulations. In the final challenge cohort, object detection dominated: five teams used the YOLO family, while other detectors included RTMDet, RF-DETR, and DETR. Four teams used segmentation approaches such as nnUNet, VM-UNet, U-Net, or DeepLabV3+ as detectors [2606.07368].

One methodological line emphasized efficient one-stage detectors coupled with pathology-specific tiling and stain handling. A YOLOv12-m submission trained only on official MIDOG 2025 datasets used 640\(\times\)640 patches, disabled scale-changing transforms and MixUp/CutMix to preserve small-object cues, enforced 50% human-derived and 50% canine-derived images per batch, and integrated multi-target Macenko normalization with 50 stochastic “stainers” during training; inference used overlapping tiling, NMS at 0.7, test-time augmentations, and Weighted Boxes Fusion, reaching \(P=0.808\), \(R=0.794\), and \(F_1=0.801\) on the preliminary test set, with runtime under 7 seconds per ROI on an NVIDIA A40 GPU [2509.02593]. A YOLOv10 ensemble submission instead operationalized data scaling and post-hoc box fusion, using five submodels, TTA with flips, WBF, and an empirically chosen 10 \(\mu\)m centroid merge to reduce telophase double counting; it achieved \(F_1=0.8407\) on the MIDOG++ test set used in that study [2509.02585].

Another line favored cascades in which a high-recall detector generated candidates and a second-stage classifier suppressed imposters. An FCOS-based submission used anchor-free candidate localization followed by a ResNet-50 classifier, with an ensemble of ResNet-50 and EfficientNet-B2 further improving candidate confirmation. On MIDOG++ validation, the FCOS baseline reached \(F_1=0.803\), while the FCOS plus ensemble-classifier cascade reached \(F_1=0.8432\), although AP decreased from 0.842 to 0.787 [2509.02597]. Team Westwood used nnUNetV2 for sensitive candidate screening and a random forest fed by three CNN backbones—EfficientNet-b3, EfficientNet-b5, and EfficientNetV2-s—for false-positive filtering, obtaining \(F_1=0.7450\) on the preliminary test set [2509.02600]. A separate two-stage framework combined an improved YOLO11x proposal generator, augmented with EMA attention and LSConv, with ConvNeXt-Tiny filtering; on a fused dataset comprising MIDOG++, MITOS\_WSI\_CCMCT, and MITOS\_WSI\_CMC, it achieved \(F_1=0.882\), improving over the single-stage improved YOLO11x baseline by \(0.035\) through precision gain from 0.762 to 0.839 at a modest recall reduction [2509.02627].

Segmentation-based formulations remained prominent because point annotations can be converted into dense supervision. A VM-UNet submission used NuClick to derive nucleus masks from MIDOG++ point annotations, trained with combined Dice and focal losses, applied Vahadane-based stain perturbation, and converted probability maps to detections via thresholding, dilation, and connected components; it reported mean leave-one-domain-out \(F_1 = 0.736 \pm 0.063\) on MIDOG++ with stain augmentation and \(F_1 = 0.754\) on the preliminary Track 1 test [2508.21033]. A teacher-student UNet submission instead formulated Track 1 as four-class segmentation—background, normal nuclei, mitosis, and hard negatives—adding contrastive representation learning and domain-adversarial training, and using a frozen teacher to generate online pseudo-masks for normal nuclei; it reached \(F_1=0.7660\), precision \(0.8264\), and recall \(0.7139\) on the preliminary test set [2509.03614]. MitoDetect++ also cast detection as segmentation with an EfficientNetV2-L U-Net and combined Jaccard, Dice, and focal losses, although that paper did not report Track 1 detection metrics on the official benchmark [2509.02586].

Not all auxiliary verification stages improved performance. One FCOS extension incorporated a Feedback Attention Ladder CNN and a fusion network for score reweighting and box correction, but the preliminary Track 1 \(F_1\) was 0.655, below the baseline FCOS value of 0.767 reported in the same work [2509.02598]. Likewise, a Faster R-CNN plus three-classifier ensemble submission achieved very high recall but critically low precision; its best official submission recorded recall \(0.9528\), precision \(0.1267\), and \(F_1=0.2237\), indicating that candidate generation alone was not the dominant bottleneck [2509.02630].

## 5. Official challenge results and empirical regularities

Eighteen teams submitted final Docker containers to Track 1. On the withheld final leaderboard, the best overall \(F_1\) score was 0.740, obtained by Lv et al. with a segmentation/discriminative “KongNet” approach; the runner-up achieved \(F_1 = 0.722\) with a YOLOv12-based method; and the best FROC-AUC was 5.347 for an SDF-YOLO submission [2606.07368]. The overall mean \(F_1\) across teams was 0.685 [2606.07368].

Performance depended strongly on ROI context. Mean \(F_1\) was 0.735 in hotspots, 0.638 in random ROIs, and 0.479 in challenging ROIs [2606.07368]. This drop was driven primarily by precision collapse rather than recall collapse: mean precision fell from 0.805 in hotspots to 0.614 in random regions and 0.400 in challenging regions, while mean recall remained comparatively stable at 0.682, 0.685, and 0.662, respectively [2606.07368]. The challenge analysis described this as false positive rates tripling in challenging ROIs, with a precision drop hotspot-to-challenging of 0.405 corresponding to a 208% increase in false-positive rate [2606.07368].

Performance also varied significantly across tumor types. Median \(F_1\) across teams ranged from 0.444 in human glioblastoma to 0.808 in canine hemangiosarcoma, and the tumor-type effect was significant by Kruskal–Wallis testing with \(p < 0.0001\) [2606.07368]. In challenging ROIs, the minimum median \(F_1\) was 0.010 for human glioblastoma, while the maximum was 0.634 for feline gastrointestinal lymphoma [2606.07368]. These findings established tumor-specific “blind spots,” especially in rare or highly pleomorphic malignancies.

Hotspot performance was not a reliable proxy for hard-context performance. The correlation between hotspot and challenging \(F_1\) across teams was weak and non-significant, with Pearson \(r = 0.36\), 95% CI \([-0.206, +0.750]\), and \(p = 0.201\) [2606.07368]. This directly challenges the assumption that success on classical mitotic hotspots is sufficient evidence of whole-slide robustness.

Post-challenge ablations further clarified which engineering choices generalized. Ensembling helped consistently: across 12 feasible ablations, mean overall \(F_1\) gain from ensembling was \(+1.549\) percentage points, with the largest mean gain in challenging ROIs at \(+2.212\) percentage points [2606.07368]. By contrast, TTA provided negligible benefit, with mean overall \(F_1\) change of \(+0.320\) percentage points and median change of \(-0.042\) percentage points [2606.07368]. This empirical result is notable because many individual submissions used TTA extensively, yet the challenge-wide analysis found it largely uninformative once measured on the final multi-context test.

## 6. Technical significance, misconceptions, and open problems

The principal lesson of MIDOG 2025 Track 1 is that hotspot-only robustness is no longer an adequate surrogate for clinical reliability. The benchmark showed that models which perform reliably in traditional hotspots can degrade sharply in random and especially challenging tissue contexts, where hard negatives and artifacts dominate the precision budget [2606.07368]. A plausible implication is that future mitosis detectors intended for whole-slide deployment will need explicit training exposure to broad tissue context, not merely centered mitosis patches.

A second lesson concerns false-positive suppression. Across the literature, high recall was often attainable, but precision remained fragile under cross-domain deployment. The Faster R-CNN submission with recall \(0.9528\) and precision \(0.1267\) is the clearest extreme [2509.02630], but the same precision–recall tension is visible in more successful systems: FCOS plus classifier refinement improved \(F_1\) while reducing AP [2509.02597], and several teams explicitly targeted necrosis, inflammation, or telophase duplicate handling as post-processing priorities [2509.02585][2509.02599]. This suggests that hard-negative modeling, threshold calibration, and biologically informed duplicate-merging heuristics can matter as much as backbone choice.

A third lesson concerns domain generalization strategy. The successful methods did not converge on a single recipe. Some emphasized multi-source training and stain-aware augmentation [2509.02593][2509.02585]; others relied on segmentation formulations with dense pseudo-supervision and domain-adversarial or contrastive regularization [2508.21033][2509.03614]. Yet the challenge paper indicates that broader contextual diversity in training data was still insufficient relative to the difficulty of the final benchmark, particularly in random and challenging ROIs [2606.07368].

Several recurring limitations remain. Many solution papers omitted exact organizer matching rules, tiling details, or post-processing thresholds, although the official challenge protocol ultimately specified Hungarian matching at \(7.5\,\mu\)m [2606.07368]. Code, seeds, and exact library versions were frequently absent [2509.02593][2509.02599]. Some sophisticated cascades failed to surpass simpler detectors, indicating that additional modules can introduce calibration problems rather than guaranteed gains [2509.02598][2509.02630].

The benchmark therefore repositions Track 1 from a relatively mature hotspot-detection problem to a harder problem of context-aware, domain-robust detection under realistic WSI heterogeneity. The official analysis recommends curating diverse hard negatives, validating on random and challenging ROIs across multiple tumors and species, favoring ensemble diversity over TTA, and optimizing thresholds with explicit false-positive budgets such as FROC rather than relying solely on hotspot \(F_1\) [2606.07368]. In that sense, MIDOG 2025 Track 1 establishes not only a leaderboard but also a methodological stress test for the next generation of mitosis detectors.

Source: https://www.emergentmind.com/topics/midog-2025-track-1