Dictionary Based Segmentation in Volumes

We present a method for supervised volumetric segmentation based on a dictionary of small cubes composed of pairs of intensity and label cubes. Intensity cubes are small image volumes where each voxel contains an image intensity. Label cubes are volumes with voxelwise probabilities for a given label. An unknown volume is segmented by cube-wise finding the most similar dictionary intensity cube. From these matches we use the probabilities from the label cubes to build a probability volume where overlapping cubes are averaged to obtain a robust label probability encoding. The dictionary is computed from labeled volumetric image data based on weighted clustering. We experimentally demonstrate our method using two data sets from material science – phantom data of solid oxide fuel cell simulations for detecting three phases, and a tomogram of a glass fiber composite used in wind turbine blades for detecting individual glass fibers.