Constrained manifold learning for the characterization of pathological deviations from normality

Abstract

This paper describes a technique to (1) learn the representation of a pathological motion pattern from a given population, and (2) compare individuals to this population. Our hypothesis is that this pattern can be modeled as a deviation from normal motion by means of non-linear embedding techniques. Each subject is represented by a 2D map of local motion abnormalities, obtained from a statistical atlas of myocardial motion built from a healthy population. The algorithm estimates a manifold from a set of patients with varying degrees of the same disease, and compares individuals to the training population using a mapping to the manifold and a distance to normality along the manifold. The approach extends recent manifold learning techniques by constraining the manifold to pass by a physiologically meaningful origin representing a normal motion pattern. Interpolation techniques using locally adjustable kernel improve the accuracy of the method. The technique is applied in the context of cardiac resynchronization therapy (CRT), focusing on a specific motion pattern of intra-ventricular dyssynchrony called septal flash (SF). We estimate the manifold from 50 CRT candidates with SF and test it on 37 CRT candidates and 21 healthy volunteers. Experiments highlight the relevance of non-linear techniques to model a pathological pattern from the training set and compare new individuals to this pattern.