The Parzen window classifier [20,27] is similar to the kNN classifier in that it is nonparametric. It obtains a nonparametric probability density estimate of the feature space for each class, allowing the class of each voxel to be determined according to the resulting posterior probabilities. In its most basic form, a voxel is assigned to the class that has the most training samples within a predetermined window of the feature space centered at the unclassified voxel. Because the Parzen window method (as well as Gaussian clustering and kNN methods) classifies voxels independently, it may yield noisy segmentations; therefore, smoothing of the originally classified data may be performed (e.g., with iterative relaxation methods) improving the appearance of the segmented image.
Was this article helpful?