Ensemble of Pathology Foundation Models for MIDOG 2025 Track 2: Atypical Mitosis Classification

Hematoxylin and eosin-stained mitotic figure (MF) counts are essential for tumor evaluation, serving both as stan-dalone and component grades in malignancy assessment (1). Mitotic figures are broadly classified into typical and atypical variants, with atypical forms--characterized by dysregulated chromatin aggregation and reflecting ge-nomic instabilities such as chromosomal instability and aneuploidy--demonstrating independent prognostic value in cancers like breast carcinoma (2, 3). However, manual enumeration and discrimination of MF variants are time-consuming and subject to substantial inter-observer variability. To address these challenges, we present a two-stage framework for automated MF classification in the MIDOG2025 Track 2 challenge (4). First, we performed parameter-efficient fine-tuning of multiple Pathology Foundation Models (PFMs) using low-rank adaptation (LoRA) (5). Training incorporated fisheye augmentation to emphasize central mi-toses (6) and Fourier Domain Adaptation (FDA) for unsupervised style transfer with ImageNet images (7). We further enhanced domain generalization by augmenting the MI-DOG2025 set with an external labeled MF dataset (8). Second, we ensembled the adapted PFMs and ConvNeXt V2 (9) to integrate complementary morphological insights into a unified classification decision (10). Our method achieved a high balanced accuracy on validation splits and also demonstrated strong performance on the Preliminary Evaluation Phase dataset, underscoring its potential for reliable, automated MF analysis.