UCS: a unified approach to cell segmentation for subcellular spatial transcriptomics

Abstract

Subcellular Spatial Transcriptomics (SST) represents an innovative technology enabling researchers to investigate gene expression at the subcellular level within tissues. To comprehend the spatial architecture of a given tissue, cell segmentation plays a crucial role in attributing the measured transcripts to individual cells. However, existing cell segmentation methods for SST datasets still face challenges in accurately distinguishing cell boundaries due to the varying characteristics of SST technologies. In this study, we propose a unified approach to cell segmentation (UCS) specifically designed for SST data obtained from diverse platforms, including 10X Xenium, NanoString CosMx, MERSCOPE, and Stereo-seq. UCS leverages deep learning techniques to achieve high accuracy in cell segmentation by integrating nuclei segmentation from nuclei staining and transcript data. Compared to current methods, UCS not only provides more precise transcript assignment to individual cells but also offers computational advantages for large-scale SST data analysis. The analysis output of UCS further supports versatile downstream analyses, such as subcellular gene classification and missing cell detection. By employing UCS, researchers gain the ability to characterize gene expression patterns at both the cellular and subcellular levels, leading to a deeper understanding of tissue architecture and function.