Foundation Model-Driven Classification of Atypical Mitotic Figures with Domain-Aware Training Strategies

TL;DR

This work tackles binary classification of normal versus atypical mitotic figures in histopathology under substantial domain shift and class imbalance. It leverages a pathology foundation model (H-optimus-0) with parameter-efficient LoRA fine-tuning, augmented by a domain-adaptive head, soft labels, hard negative mining, adaptive focal loss, and metric learning via supervised contrastive loss; a domain-aware sampling strategy further biases training toward minority and harder examples. The approach demonstrates strong cross-domain generalization with a mean balanced accuracy of $BA=0.851\pm0.037$ across 10 domains, though substantial OOD variability remains, highlighting both the promise and challenges of foundation-model-based domain adaptation in AMF/NMF classification. Overall, the work provides a practical framework for domain-aware adaptation of pathology foundation models that improves minority-class detection while maintaining reasonable majority-class performance, with implications for robust Mitotic Figure analysis across diverse clinical settings.

Abstract

We present a solution for the MIDOG 2025 Challenge Track~2, addressing binary classification of normal mitotic figures (NMFs) versus atypical mitotic figures (AMFs). The approach leverages pathology-specific foundation model H-optimus-0, selected based on recent cross-domain generalization benchmarks and our empirical testing, with Low-Rank Adaptation (LoRA) fine-tuning and MixUp augmentation. Implementation includes soft labels based on multi-expert consensus, hard negative mining, and adaptive focal loss, metric learning and domain adaptation. The method demonstrates both the promise and challenges of applying foundation models to this complex classification task, achieving reasonable performance in the preliminary evaluation phase.