Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

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

Created by
  • Haebom

Author

Mieko Ochi, Bae Yuan

Outline

This paper classifies mitotic figures into typical and atypical types. The number of atypical mitotic figures strongly correlates with tumor aggressiveness. Accurate classification is therefore essential for predicting patient prognosis and allocating resources, but it remains a challenging task even for expert pathologists. This study utilized pathology-based models (PFMs) pretrained on a large-scale histopathology dataset to perform parameter-efficient fine-tuning via low-dimensional adaptation. Furthermore, we integrated ConvNeXt V2, a state-of-the-art convolutional neural network architecture, to complement the PFMs. During training, we used the fisheye transform to highlight mitoses and applied Fourier domain adaptation using ImageNet target images. Finally, we ensemble multiple PFMs to integrate complementary morphological insights, achieving competitively balanced accuracy on a preliminary evaluation dataset.

Takeaways, Limitations

•
Takeaways:
◦
Combining pre-trained PFMs with a state-of-the-art CNN architecture (ConvNeXt V2) on a large dataset to improve mitotic classification performance.
◦
Parameter-efficient fine-tuning and performance improvement through techniques such as low-dimensional adaptation and Fourier domain adaptation.
◦
Integrating diverse morphological information and improving performance through ensemble techniques.
◦
Achieving competitive results on preliminary evaluation datasets.
•
Limitations:
◦
Only results for the preliminary evaluation phase dataset are presented, so further verification of generalization performance is needed.
◦
Lack of detailed description of the size and composition of the dataset used.
◦
Lack of comparative analysis with other similar studies.
◦
Further research is needed for clinical application.
👍
Made with Slashpage