Overview / Usage

• Rectangle patches that contains the whole tumor region (plus 10% surrounding region)
  is extracted from each modality and view image
• All the patches are resized to the same size (224*224)
  Each patient, 4 images are fed into GoogLeNet, and get 4096 features (1024
  for each image) at the last layer
• Random forest (RF) is implemented to select important features from all
  4096 features (based on Gini impurity)
• 137 features are kept and fed into another RF for final classification into
  Benign (-) or Malignant (+)