Figure 2. Schematics of the DSIFT-BOF framework.

A codebook feature space is created by extracting DSIFT descriptors at fixed grid locations from all the images that are dedicated to vocabulary building. A codebook is generated by running a clustering algorithm that partitions the codebook feature space into k regions. The centroids of these regions (a.k.a. clusters) represent the codebook terms. The codebook allows for BoF representations, term vectors, to be assigned to training and test images. The term vector is a histogram that indicates the number of occurrences of the codebook terms in an image. Term vectors are assigned to the training images, and will be further on used as ground truth data by a classifier. An image is tested by assigning a term vector to it, and running this term vector through a classifier that uses the term vectors of the training images to indicate its fibrosis stage. The classifier that we use in this experiment is weighted k-NN.