The nearest-neighbor technique requires certain voxels to be defined by the operator and then used as references for automated classification of new data. Following the supervised training of the data in which a few characteristic voxels are sampled and labeled manually the algorithm assigns each unclassified voxel to the class of its closest neighbor in the feature space. The feature space can be made of parameters or any function of the images such as gradient information or texture, although the image intensities are typically used to form the feature space. For example, in double-echo MR images, two image intensities are acquired at each voxel, thereby forming a two-dimensional feature space. The k-nearest-neighbor (kNN) classifier [20,27] extends the pre ceding concept to the k nearest neighbors among the class samples, classifying new voxels according to the majority vote of the k closest neighbors in the training set. Avoxel may be left unclassified if the nearest neighbor in the training set is further than some predetermined distance.
Was this article helpful?