# PathoDuet: Foundation Models for Pathological Slide Analysis of H&E and IHC Stains

**Authors:** Shengyi Hua, Fang Yan, Tianle Shen, Xiaofan Zhang  
**arXiv:** [2312.09894](https://arxiv.org/abs/2312.09894)  
**Date:** 2023-12-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2312.09894v1/

## Key points

- Authors introduce PathoDuet, a novel approach to developing pathological foundation models using self-supervised learning methods on digitized histopathological data
- PathoDuet consists of pretrained models specifically designed for histopathological images
- New self-supervised learning framework includes pretext token and task raisers to leverage relationships between images like multiple magnifications and stains
- Two pretext tasks, cross-scale positioning and cross-stain transferring, are introduced for pretraining the model on H&E images and transferring it to IHC images
- PathoDuet models are validated through extensive evaluation across various downstream tasks such as colorectal cancer subtyping, WSI-level classification in H&E field, expression level prediction of IHC markers, and tumor identification in IHC field
- Experimental results demonstrate the superiority of PathoDuet models over existing methods for most tasks
- Availability of codes and models on GitHub provides valuable resources for researchers to further explore and utilize these innovative approaches

## Summary

In their paper titled "PathoDuet: Foundation Models for Pathological Slide Analysis of H&E and IHC Stains," authors Shengyi Hua, Fang Yan, Tianle Shen, and Xiaofan Zhang introduce a novel approach to developing pathological foundation models using self-supervised learning methods on digitized histopathological data. These  are pretrained to serve as a strong basis for various  in the field of . However, the direct application of existing methods is hindered by the significant gap between natural images and histopathological images. To address this challenge, the authors present PathoDuet, a series of pretrained models specifically designed for histopathological images. They propose a new self-supervised learning framework that incorporates a pretext token and task raisers to leverage relationships between images such as multiple magnifications and stains. Two pretext tasks, cross-scale positioning and cross-stain transferring, are introduced to pretrain the model on Hematoxylin and Eosin (H&E) images and transfer it to immunohistochemistry (IHC) images. The efficacy of PathoDuet models is validated through extensive evaluation across a range of downstream tasks. These tasks include patch-level colorectal cancer subtyping, whole slide image (WSI)-level classification in H&E field, expression level prediction of IHC markers, and tumor identification in IHC field. The experimental results demonstrate the superiority of PathoDuet models over existing methods for most tasks, highlighting the effectiveness of the proposed pretext tasks. Overall,represents a significant advancement in leveraging self-supervised learning methods for developing foundation models in histopathology. The availability of codes and models on GitHub provides researchers with valuable resources to further explore and utilize these innovative approaches in their own work.