Extended Data Fig. 3: Benchmarking UTAG and competing methods against expert labels.
From: Unsupervised discovery of tissue architecture in multiplexed imaging
a) Results of each method on healthy lung data to segment microanatomical domains. Number of latent topics for SpaGene was set to 10 to capture the diverse target phenotypes. Due to supporting only single images, SpaGene topics were relabeled using agglomerative clustering to consistently label topics across slides. b) Results of each method on tumor vs. stroma on upper tract urothelial carcinoma. Number of latent topics for SpaGene was set to four to differentiate tumor versus stroma. c) Example of running UTAG, SpatialLDA, and SpaGene to demonstrate the difference in performance. The color mapping in this panel is different for each method as all three methods are unsupervised. d) Same as c) but with domain colors remapped to correspond to the ones from expert labels for ease of visual comparison. For a) and b), Data in boxplots are presented by minimum, 25th percentile, median, 75th percentile, and maximum. Values outside of 1.5 times interquartile range are classified as outliers and are denoted as fliers.
