Strong links exist between mechanical dyssynchrony and the response to cardiac resynchronization therapy (CRT). Recent publications recommend identifying correctable dyssynchrony patterns with a specific motion and deformation signature. The learning of these patterns is visual and highly subjective. We take advantage of statistical atlas and dimensionality reduction tools to learn a representation of these patterns. We hypothesize that myocardial motion patterns belong or lie close to a nonlinear manifold, and model them as a pathological deviation from normality. Furthermore, we propose distances to compare new subjects with those patterns and with normality. We evaluate the value of this approach on 2D echocardiographic sequences from CRT candidates at baseline, with pacing on, and at 1-year follow-up. We demonstrate that relating pattern changes with patient response is valuable, and paves the way towards better therapy planning.
Duchateau N, Piella G, Frangi A, De Craene M.. Learning pathological deviations from a normal pattern of myocardial motion: Added value for CRT studies?. In: -. Machine Learning and Medical Imaging 1st Edition. 1 ed. Academic Press, Elsevier; 2016. p. 365-382.