Comparison of cell-type inference between Visium HD and Xenium Spatial Transcriptomics Technologies

Spatial transcriptomics (ST) technologies enable simultaneous profiling of gene expression and spatial organisation at single-cell resolution. A critical step in ST data analysis is the automated assignment of cell-type identities to spatially profiled cells, a task for which dedicated methods remain scarce and tools developed for single-cell RNA sequencing are routinely repurposed despite the distinct characteristics of spatial data. This thesis presents a comparative analysis of cell-type inference across two complementary ST platforms: Visium HD, a sequencing-based whole-transcriptome technology, and Xenium In Situ, an imaging based targeted-panel platform, applied to adjacent colorectal cancer tissue sections, with a matched non-spatial Chromium single-cell RNA-seq dataset as annotation reference. SingleR was selected following evaluation of four candidate tools. Shannon entropy analysis revealed that high annotation uncertainty is a pervasive feature of spatial data driven primarily by gene panel size: at matched gene panel size, Visium HD and Xenium produced statistically indistinguishable entropy distributions. A tile-based spatial comparison demonstrated high cross-technology concordance at the ~106μm scale, with a global Pearson correlation of r = 0.947 and a dominant cell-type agreement of 87.5%. These findings suggest that gene panel design is a more critical determinant of annotation reliability than the choice of detection technology.