Segmentation in Feature Space

Selection of an image acquisition technique is intentional in medical imaging. It can be assumed that pixel or voxel values in a medical image cover more semantics with respect to object class membership than intensity in a photograph. Hence, image segmentation can be done as classification in feature space where image intensities are the features. The dimensionality of feature space is usually low, and the number of samples characterizing object classes is high. Typical classifiers discussed in this chapter take this into account and estimate likelihood functions from samples. Classification is then done by computing a posteriori probabilities for each object class. Clustering in feature space will be discussed as well. Without requiring training, clustering may directly lead to segmentation. Even if this is not the case, clustering may be used to reduce the work load for producing the training data.