Mitosis Detection in the Wild: Multi-Tumor and Context-Aware Generalization in the MIDOG 2025 Challenge

The MIDOG 2025 challenge evaluated automated mitosis detection across 365 diverse tumor cases spanning 12 different human, canine, and feline types to assess real-world clinical applicability. Results showed top F1 scores of 0.740 for detection and 0.908 balanced accuracy for atypical mitotic figure classification, but revealed significant performance degradation in challenging tissue areas where false positives tripled, highlighting major limitations in current AI architectures.

The MIDOG 2025 challenge represents a critical evolution in computational pathology benchmarking by moving beyond controlled laboratory conditions to evaluate genuine clinical robustness. Traditional mitosis detection benchmarks focused narrowly on scanner-induced domain shifts and hand-curated hotspot regions, creating an artificial evaluation environment disconnected from actual pathology workflows. This challenge fundamentally changes that paradigm by requiring models to perform across random tissue areas and challenging regions rich in hard negatives—scenarios representative of whole-slide scanning in clinical practice.

The results expose significant architectural limitations in current state-of-the-art models. While best-performing submissions achieved respectable F1 scores of 0.740 in hotspot regions, performance degraded substantially in challenging areas, with false positive rates tripling. Equally concerning, models demonstrated substantial variance across tumor types, suggesting blind spots when encountering rare or morphologically complex malignancies. This heterogeneous performance across biological contexts rather than just technical parameters reveals that generalization remains the core unsolved problem in digital pathology.

For the computational pathology industry, these findings validate that laboratory benchmarks provide false confidence in clinical readiness. Researchers and vendors developing diagnostic AI must now prioritize robustness across biological diversity rather than optimizing for established test sets. The modest improvements from ensembling (1.5 percentage points) and negligible gains from test-time augmentation indicate that incremental methodological refinements yield diminishing returns without addressing fundamental architectural limitations.

Moving forward, developers should prioritize training strategies that explicitly target morphological diversity and incorporate mechanisms for handling hard negatives in realistic tissue contexts. The gap between hotspot and wild performance metrics will increasingly define competitive advantage in clinical deployment.

• →Top-performing mitosis detection models achieved F1 scores of 0.740 but showed 3x higher false positive rates in challenging tissue regions versus hotspots.
• →Performance varied significantly across 12 tumor types, revealing that current AI architectures have blind spots with rare or pleomorphic malignancies.
• →Ensembling provided only marginal improvements (1.5 percentage points in F1), while test-time augmentation showed no meaningful benefit.
• →Clinical-grade mitosis detection requires evaluation across diverse biological contexts, not just scanner-induced domain shifts in curated hotspots.
• →The transition from hotspot-only benchmarks to multi-contextual evaluation fundamentally redefines readiness for real-world pathology deployment.