Overcoming Domain Shift in Atypical Mitotic Figure Detection with Deep Ensemble Learning

Area of Science:

• Computational pathology
• Histopathology image analysis
• Machine learning in oncology

Background:

• Accurate classification of atypical mitotic figures (AMFs) is crucial for cancer prognosis.
• Current deep learning models struggle with generalization in diverse histopathological settings.
• Developing robust AMF detection pipelines is essential for clinical translation.

Purpose of the Study:

• To present a robust and reproducible pipeline for the detection of atypical mitotic figures (AMFs).
• To evaluate the generalization capabilities of the developed pipeline across varied datasets and clinical conditions.

Main Methods:

• Compiled a large dataset from three public histopathology image sources.
• Trained an ensemble of three ConvNeXt models using a 3-fold cross-validation bagging strategy.
• Validated the pipeline on an internal hold-out set and the MICCAI MIDOG2025 Challenge.

Main Results:

• Achieved 89.18% balanced accuracy on an internal hold-out dataset.
• Demonstrated excellent generalization in the MICCAI MIDOG2025 Challenge with 88.94% balanced accuracy, ranking #8.
• Minimal performance drop indicates robustness against variations in tissue, staining, and scanners.

Conclusions:

• The developed pipeline offers a validated, foundational tool for clinical AMF analysis.
• The robust generalization performance supports its applicability in diverse real-world histopathology scenarios.
• This work addresses the critical need for reliable automated AMF detection in cancer diagnostics.