Fig. 4: Tissue structure discovery by ROSIE.
From: ROSIE: AI generation of multiplex immunofluorescence staining from histopathology images
Discovery of tissue structures using biomarkers generated by ROSIE on the Stanford-PGC test dataset. Five tissue structures are identified using a graph partitioning algorithm that clusters cells based on their expression profiles and neighboring cells. This algorithm is performed on both the ground truth measured and ROSIE-generated biomarker expressions and then reconciled to a common label set. A Visualizes several representative samples of tissue structures discovered using the ground truth CODEX measurements, ROSIE-generated expressions, and the morphology baseline method. B Left: We report the F1 score (N = 635,649 cells) by comparing the structures discovered using ground truth, ROSIE-generated biomarkers, and morphology features. Data are presented as mean values with error bars as the 95% bootstrapped confidence intervals. Right: ARI score is also reported by comparing the unlabeled discovered clusters, where each dot is a sample. Box plots show the median (center line), 25th and 75th percentiles (box edges), and the minimum and maximum values.
