Lung Segmentation Using Incremental Sparse NMF

The traditional way to model the visual appearance of the image is to define the spatial interactions of the image voxels in terms of their neighboring voxels. A new spatial interaction model was developed for the 3D lung data by extracting new spatial features based on NMF. Let GNx,y,z ∈ QIx×Iy×Iz be the image signals of the neighborhood of the voxel (x, y, z). By including the image signals of the neighborhood of all voxels, a 4D matrix G ∈ Q Z×Ix×Iy×Iz is composed. To process the 3D lung data using NMF, GNx,y,z of each voxel is represented as a vector gNx,y,z in the input data matrix A ∈ QIxIyIz×XY Z , as shown in Fig. 2. Using conventional NMF, A is decomposed as follow: A ≈WH (1)