We address the problem of extracting meaningful, uncorrelated biological modes of variation from tangent space shape coordinates in 2D and 3D using non-Euclidean metrics. We adapt the maximum autocorrelation factor analysis and the minimum noise fraction transform to shape decomposition. Furthermore, we study metrics based on repated annotations of a training set. We define a way of assessing the correlation between landmarks contrary to landmark coordinates. Finally, we apply the proposed methods to a 2D data set consisting of outlines of lungs and a 3D/(4D) data set consisting of sets of mandible surfaces. In the latter case the end goal is to construct a model for growth prediction and simulation.
|Title of host publication||Medical Image Computing and Computer-Assisted Intervention - MICCAI 2002, 5th Int. Conference, Tokyo, Japan|
|Publication status||Published - 2002|