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

2026-06-05 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition, Health & Medical Research · Depth: Advanced, quick

Summary

The MItosis DOmain Generalization (MIDOG) 2025 challenge evaluated automated mitosis detection algorithms for robustness across diverse biological and contextual variances, moving beyond scanner-induced domain shifts. The challenge utilized a test dataset of 365 cases, encompassing 12 distinct human, canine, and feline tumor types digitized across multiple scanning platforms. Unlike previous benchmarks, it required detection in random tissue areas and challenging regions rich in hard negatives, not just hand-selected hotspots. A second track focused on classifying atypical mitotic figures (AMFs). Results from 18 teams in the detection track showed F1 scores up to 0.740, while 21 AMF submissions achieved balanced accuracy up to 0.908. Analysis revealed significant performance degradation in challenging ROIs, with false positive rates tripling, and varied performance across tumor types, indicating "blind spots" in existing architectures. Ensembling improved F1 by 1.5 percentage points and balanced accuracy by 1.3 percentage points, whereas Test-Time Augmentation (TTA) offered no relevant improvement.

Key takeaway

For computational pathologists developing automated mitosis detection systems, recognize that current models often fail in "in the wild" clinical scenarios. Your systems will likely show significant performance degradation in challenging tissue regions and exhibit "blind spots" for rare or pleomorphic malignancies. Prioritize developing architectures that generalize robustly across diverse biological and contextual variances. Consider implementing ensembling techniques, which improved F1 scores by 1.5 percentage points, and rigorously test your models on multi-contextual datasets beyond traditional hotspots to ensure clinical reliability.

Key insights

"In the wild" mitosis detection demands models robust to diverse biological and contextual variance.

Principles

• Current architectures exhibit "blind spots" for rare malignancies.
• Performance significantly degrades in challenging tissue regions.
• Ensembling improves detection robustness across diverse contexts.

Method

Evaluate mitosis detection algorithms using multi-species, multi-scanner datasets, assessing performance in random tissue areas and challenging regions beyond traditional hotspots.

In practice

• Test models on diverse tumor types, including rare ones.
• Incorporate challenging ROIs into model evaluation.
• Apply ensembling to enhance detection reliability.

Topics

• Automated Mitosis Detection
• Computational Pathology
• Domain Generalization
• MIDOG 2025 Challenge
• Atypical Mitotic Figures
• Model Robustness

Best for: AI Scientist, Research Scientist, Computer Vision Engineer

Related on AIssential

• Mitosis Detection in the Wild: Multi-Tumor and Context-Aware Generalization in the MIDOG 2025 Challenge — Automated mitosis detection models exhibit significant generalization gaps "in the wild" beyond curated hotspot regions.
• Non-invasive profiling of the tumour microenvironment with spatial ecotypes — A new machine learning framework enables non-invasive profiling of tumor microenvironment spatial ecotypes from liquid biopsies, predicting immunotherapy response.
• Meta-encoder: a unified integration framework for multiple pathological foundation models in cancer detection — The Meta-Encoder integrates multiple pathological foundation models to improve cancer detection and molecular characterization.
• GigaTIME: Scaling tumor microenvironment modeling using virtual population generated by multimodal AI — GigaTIME uses multimodal AI to convert common H&E slides into virtual mIF images, enabling population-scale precision immuno-oncology research.
• Margin-Consistent Deep Subtyping of Invasive Lung Adenocarcinoma via Perturbation Fidelity in Whole-Slide Image Analysis — Margin consistency with Perturbation Fidelity enhances deep learning robustness and accuracy for lung adenocarcinoma subtyping.
• The Dark Regulome: Disentangling Predictability from Regulation in Genomic Foundation Models — A diagnostic separates sequence predictability from true regulatory influence in genomic foundation models, revealing distinct functional layers.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.