Figure 2
Feature-guided H&E sample selection and virtual IF staining with SHIFT. (a) Distribution of the 16 latent features (L1-L16) extracted by VAE from sample H&E tiles. (b) t-SNE embedding of latent feature representations of sample H&E tiles, faceted by sample identity. Each point in each plot represents a single H&E tile. Contour lines indicate point density. (c) SHIFT model test performance for optimal (B and D) and non-optimal (A and B) training set sample compositions. The paired H&E and IF images from samples B and D were subdivided into smaller images B = {B1,B2} and D = {D1,D2,D3,D4,D5} to avoid regions of IF that exhibited substantial autofluorescence. The x-axis labels indicate sample identity, where each letter corresponds to a unique sample and each number corresponds to a subset of that sample. Each n denotes the number of image tiles that were extracted from that sample. Plots for sample subsets are not show if that sample subset was a component of a model’s training set. *p < .05; for three group comparisons we used the Friedman test with Nemenyi post-hoc test; for two group comparisons we used the Wilcoxon signed-rank test. White dots in violin plots represent distributional medians. (d) Visual comparison of virtual staining methods. The ensemble results are attained by averaging the output images of SHIFT and Label-Free Determination (LFD) models. See also Supplementary Fig. S2. (e) Test performance comparison of virtual staining methods. The x-axis labels indicate sample identity, where each letter corresponds to a unique sample and each number corresponds to a subset of that sample. Each n denotes the number of image tiles that were extracted from that sample. Plots for sample subsets B1 and D5 are not show because those sample subsets were components of the models’ training sets. *p < .05; Friedman test with Nemenyi post-hoc test. White dots in violin plots represent distributional medians.
