Cluster Based Vector Attribute Filtering

Abstract

Abstract Morphological attribute filters operate on images based on properties or attributes of connected components. Until recently, attribute filtering was based on a single global threshold on a scalar property to remove or retain objects. A single threshold struggles in case no single property or attribute value has a suitable, usually multi-modal, distribution. Vector-attribute filtering allows better description of characteristic features for 2D images. In this paper, we apply vector-attribute filtering to 3D and incorporate unsupervised pattern recognition, where connected components are classified based on the similarity of feature vectors. Using a single attribute allows multi-thresholding for attribute filters where more than two classes of structures of interest can be selected. In vector-attribute filters automatic clustering avoids the need for either setting very many attribute thresholds, or finding suitable class prototypes in 3D and setting a dissimilarity threshold. Explorative visualization reduces to visualizing and selecting relevant clusters. We show that the performance of these new filters is better than those of regular attribute filters in enhancement of objects in medical images.