An AI-based cell segmentation pipeline for studying immune markers from brightfield microscopy images

Abstract

Background: Cell segmentation is a essential task in analysing the tumormicroenvironment (TME), which encompasses vascular cells and cells of theinnate and adaptive immune system in and around the tumor. The impact of theTME on survival and effectiveness of immune therapies has been shown fordifferent tumor types e.g. gliomas. Methods: Here, we developed an AI-based pipeline to analyse the TME inbrightfield microscopy images created by immunohistochemistry (IHC) applying CD3, CD4, CD8, CD20, CD163 and Iba1 antibodies. The pipeline consist offiltering the tumor areas by a residual neural network with 34 layers andsegmenting the immune and non-immune cells by a U-net based model. Results: For most of the cell types, a filtering classifier trained on all cell typeshad a better performance than a classifier trained on one cell type. For the cellsegmentation, the segmentation classifier used in the pipeline outperformed thestate-of-the-art method Cellpose and Watershed as classical approach. Conclusions: This AI-based cell segmentation pipeline can be used to analysethe TME of a big set of whole slide images without manual effort and toinvestigate the relationship between immune cells and survival in several tumortypes. One slide can be analysed in about 30 minutes, thus the pipeline has thepotential to be used in routine clinical diagnosis and could provide informationfor therapy suggestions.