TLDR: Researchers Mieko Ochi and Yuan Bae developed a two-stage AI framework for accurate atypical mitosis classification, crucial for tumor prognosis. The method involves fine-tuning Pathology Foundation Models (PFMs) using Low-Rank Adaptation (LoRA), enhanced with fisheye augmentation to emphasize mitotic figures and Fourier Domain Adaptation (FDA) for style transfer. By ensembling multiple PFMs (UNI, Virchow, Virchow2), the framework integrates diverse morphological insights, achieving high balanced accuracy and demonstrating potential for reliable, automated pathology analysis to reduce pathologist workload.
Accurately identifying atypical mitotic figures in tissue samples is crucial for diagnosing and predicting the aggressiveness of tumors, yet it remains a complex task even for experienced pathologists. These figures, which indicate genomic instabilities, have significant prognostic value in cancers like breast carcinoma. However, manual counting and differentiation are time-consuming and prone to inconsistencies between observers.
To address these challenges, researchers Mieko Ochi and Yuan Bae from the Department of Pathology, Japanese Red Cross Medical Center, Japan, have developed a sophisticated two-stage artificial intelligence framework. This innovative approach aims to automate and improve the accuracy of atypical mitosis classification, as detailed in their paper, Ensemble of Pathology Foundation Models for MIDOG 2025 Track 2: Atypical Mitosis Classification.
A Two-Stage AI Framework for Enhanced Accuracy
The first stage of their framework involves fine-tuning multiple Pathology Foundation Models (PFMs). These are powerful AI models pre-trained on vast datasets of histopathology images. The fine-tuning process uses a technique called Low-Rank Adaptation (LoRA), which efficiently adapts these large models to the specific task of classifying mitotic figures without needing to retrain the entire model.
To further enhance the models’ performance and generalization, the training incorporates several key augmentations:
- Fisheye Augmentation: This optical distortion technique emphasizes the central region of an image, where mitotic figures are typically located, helping the models focus on these critical areas.
- Fourier Domain Adaptation (FDA): This method performs unsupervised style transfer using generic ImageNet images. It helps mitigate domain shifts that can arise from differences in scanner types and staining protocols across various institutions, making the models more robust.
- External Labeled Datasets: The MIDOG2025 dataset is augmented with additional labeled mitotic figure datasets to provide a richer and more diverse training experience.
In the second stage, the adapted PFMs are combined using an ensemble learning approach. The researchers selected three prominent PFMs—UNI, Virchow, and Virchow2—known for their strong performance in pathology tasks. Each of these models has been pre-trained on different histopathology datasets, leading to diverse feature extraction capabilities. By ensembling their predictions, the framework integrates these complementary insights into a unified, more accurate classification decision.
Also Read:
- Enhancing Mitosis Detection in Digital Pathology with a Hybrid AI Approach
- Adaptive AI Strategies Advance Atypical Mitosis Classification in Pathology
Key Findings and Impact
The research included an ablation study to evaluate the effectiveness of fisheye and FDA augmentations. It was found that fisheye augmentation alone yielded the highest balanced accuracy for the UNI model, while the combination of both also performed well. More significantly, the ensemble of all three PFMs (UNI + Virchow + Virchow2) demonstrated a substantial improvement in balanced accuracy, increasing it by approximately 5% compared to the best single model (Virchow2).
The submitted ensemble models for the Preliminary Evaluation Phase showed promising results. While the combination of fisheye and FDA achieved a higher overall balanced accuracy, the fisheye-only model exhibited more stable performance across different domains, highlighting the importance of understanding how augmentations affect generalization.
This framework represents a significant step forward in automated histopathological image analysis. By leveraging the power of foundation models and advanced augmentation and ensemble techniques, it offers a reliable method for classifying atypical mitoses. This has the potential to reduce the workload on pathologists and improve the consistency and accuracy of tumor evaluation in clinical practice.
