Towards deformable shape modeling of the left atrium using non-rigid coherent point drift registration

Abstract

Modeling the deformable shape of the left atrium is of strong interest for many applications in cardiac diagnosis and intervention. In this paper, we propose a method for left atrium shape modeling using non-rigid point cloud registration. In particular, we build upon the concept of Coherent Point Drift (CPD) registration that considers the alignment as a probability density estimation problem. Based on the set of non-rigidly registered point clouds, we perform a principle component analysis to establish a deformable shape model. In an experimental study on ten clinical data sets, we evaluated the registration accuracy in terms of average mesh-to-mesh distance, as well as on anatomical landmarks on the left atrium. With the proposed method, we achieved registration results with an average mesh-to-mesh error of 3.4mm. The average landmark offset was 8.5mm.