Novel method to automatically identify medial node correspondences between two images

Abstract

Many modern forms of segmentation and registration require manual input, making them tedious and time-consuming processes. There have been some successes with automating these methods, but these tend to be unreliable due to inherent variations in anatomical shapes and image quality. It is toward this goal that we have developed methods of identifying correspondences in two images between medial nodes; image features related to anatomical structures. Medial based image features are used because they have proven robust against image noise and shape variation, and provide rotationally invariant properties of dimensionality and scale, while preserving orientation information independently. We have introduced several novel metrics for comparing the medial and geometric relationships between medial nodes and different cliques of medial nodes (a clique is a set of multiple medial nodes). These metrics overcome problems introduced by symmetry between cliques and provide increasing discriminability with the size of the clique. In this paper, we demonstrate medial-based correspondences and validate their specificity with standard Receiver Operator Characteristic (ROC) analysis. It is believed that our method of locating corresponding medial features may be useful for automatically locating anatomical structures or generating landmarks for registration.