Fig. 2: Slide pre-processing, classifier performance, and pattern discovery. | npj Artificial Intelligence

Fig. 2: Slide pre-processing, classifier performance, and pattern discovery.

From: A multimodal framework to identify molecular mechanisms driving patient group-associated morphology through the integration of spatial transcriptomics and whole slide imaging

Fig. 2

a Example of tiles cut from annotated invasive ROIs. b UMAP of RetCCL features before and after AI-FFPE style transfer on frozen spatial dataset tiles. c Pathologist tile classification performance and example tiles for QC. d Comparison of attention-MIL average AUC across 3-fold CV, without and with self-attention function before model training. e UMAP representation of high-attention features colored by leiden cluster, and all features colored by leiden clusters from separate high and low attention clustering. f Example tile pattern assignments in high attention regions.

Back to article page