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

The morphological classification of atypical mitotic figures (AMFs) is a critical prognostic task in histopathology, but deep learning models often lack generalization across diverse clinical settings. This study presents a robust and reproducible pipeline for AMF detection. We compiled a large dataset from three public sources and trained an ensemble of three ConvNeXt models using a 3-fold cross-validation based bagging strategy. The pipeline achieved a balanced accuracy of 89.18% on an internal hold-out set and demonstrated excellent generalization in the MICCAI MIDOG2025 Challenge with a comparable 88.94% balanced accuracy, securing rank #8 on the challenge leaderboard. The minimal performance drop confirms the detection robustness against variations in tissue types, staining, and scanners, providing a validated foundational tool for clinical AMF analysis.