A Foundation Model for Cell Segmentation from Immunohistochemistry: A Pilot Study

Cell segmentation plays an important role in the basic research and disease diagnosis of immunohistochemistry (IHC). Although frontiers of large foundation models have been well investigated in H&E stained computational pathology, few works have focused on the diverse and complex analytics of IHC cells, meaning there is currently no foundation model specifically pre-trained on IHC datasets. Furthermore, existing foundation models fail to adequately consider the diversity of contextual priors in IHC images, such as imaging, preparation, and staining methods. Additionally, in downstream segmentation tasks, the popular transformer architecture lacks designs for capturing the subtle structural features of cells. Therefore, in this paper, we establish the first generalizable IHC foundation model (IHCFM), which is pre-trained on a large-scale IHC dataset consisting of 347,226 patches across different hospitals. Meanwhile, to address the diversity of IHC data, we introduce a method to decouple image contextual priors on modeling. Then, we propose the downstream segmentation network, which can effectively capture global and local information in IHC cells. Finally, we evaluate the segmentation performance of the model on a large-scale dataset with 1.2 million manually annotated cells.