Figure 1.

Workflow for the unsupervised learning using a bag-of-words paradigm. In step (1) the cortex part of the biopsy sample was used; (2) the Reinhard stain color normalization method applied; (3) each biopsy sample image was tiled into 256 × 256 pixel patches; (4) we extracted features from each patch using the transfer learning method in deep learning; (5) unsupervised machine learning algorithms called K-means clustering was applied; and finally (6) a histogram representation for each biopsy sample was created to describe the distribution of each type of cluster at the patient level.