Figure 3

Color based segmentation and k-means clustering of grid cells into epithelial and non-epithelial regions based on color area ratios. A representative micrograph of prostate tissue section at 20× (A) is segmented into four regions based on the pixel colors of (B) white, (C) light pink, (D) dark pink and (E) purple, using a k-means clustering algorithm. (F) The epithelial areas of the training image were marked by an experienced prostate cancer researcher and were arranged into 20 × 20 pixel grid cells. (G) The four color ratios were calculated in each grid cell. Knowing the epithelial and non-epithelial regions in training sets, we classify the grid cells into two clusters. A scatter plot shows these clusters in the space of three colors, which have small overlap.