Weakly-Supervised Cell Segmentation for Multiplex Immunohistochemistry Images

Anonymous

TL;DR: Cell segmentation of mIHC with dot annotations by exploiting superpixels and multi-resolution supervision loss.

Abstract: Multiplex immunohistochemistry (mIHC) is a novel scalable method of staining multiple cell types in a single tissue slice. In this paper, we propose a new method to automatically segment multiple cell types from a mIHC whole slide image. Our method only requires domain experts to provide a limited number of weak annotations (i.e., labeled dots placed at the centers of cells), while still achieving high quality segmentation. In particular, we (1) expand dot labels to mask annotations via superpixels; (2) introduce a multi-resolution supervision loss; and (3) leverage color deconvolution networks to further refine segmentation boundaries. Empirical evaluation on pancreatic cancer tissue slides demonstrates the efficacy of our method in providing an unprecedented amount of data from a single tissue section. Combining mIHC and the cell segmentation methods described herein would enable large scale studies of the immune contexture of cancer with minimal annotation effort from domain experts.

Keywords: Cell Segmentation, Multiplex Immunohistochemistry