Figure 2

Examples of pseudo-color outputs of Aperio image analysis algorithms from which the predictive features were calculated. (a) Analysis square from an H&E WSI in the training set (75% overlap with the slide-level annotated cancer outlined in yellow). (b) Analysis square-level annotation of (a), with benign epithelium in green, malignant epithelium in red, gland lumens in white, and stroma in blue. The percentage of malignant epithelium is used as the ground truth for training. (c–e) Outputs of the PPC algorithms. Positive pixels are in red/orange/yellow, and negative pixels in blue. The percentages of positive pixels is taken to be the percentages of the analysis square occupied by nuclei, cytoplasm, or stroma. (f) Analysis square corresponding to (a) taken from the corresponding IHC WSI. (g,h) Outputs of the CD algorithms. Positive pixels are in yellow/orange/red, and negative pixels in blue. (i) Output of the CE algorithm. Positive pixels for red and brown components are in green-cyan and red-purple, respectively. (j) Table of features derived from the outputs of the image analysis algorithms.