Fig. 2: Computational efficiency of Sopa in terms of RAM and time on different dataset sizes.

a Cellpose segmentation comparison: with and without patching. The RAM usage is given per core. b Baysor segmentation comparison: with and without patching. The RAM usage is given per core. c Examples of cell boundaries before resolving the conflicts over overlapping patches when running Cellpose segmentation on DAPI staining (MERSCOPE human liver hepatocellular carcinoma dataset). On overlapping regions, cells are segmented twice (middle and right). For each conflict, their IOMA determines whether or not to merge the two cell boundaries. d UMAP showing the difference between the resolution with and without the patching process. e Violin plots showing the intersection-over-min-area density of segmentation conflicts when using patches (for both Cellpose and Baysor). When resolving a conflict, the two good cases are either (i) a high concordance between the two cells (which will be merged), or (ii) a low concordance between them (the two cells are kept). IOMA below 0.07 or above 0.8 correspond to good conflict resolution cases. f Channels averaging for each cell: Sopa and standard average inside numpy masks. g Counting each gene inside each cell: with Sopa compared to GeoPandas join operation on the whole DataFrame. h Writing image as a tiff file for the Xenium Explorer: with Sopa compared to what is recommended by 10× Genomics, i.e., loading the whole image in memory. Source data are provided as a Source Data file.