Application of Laplacian Surface Deformation and self-organizing maps to calculate shape correspondence for statistical shape modeling

This paper presents a pairwise approach of finding shape correspondence for the construction of SSM, applied to distal femur bone. The correspondence was calculated by adapting a representative shape (template mesh) onto the training shapes in two steps. The first step was global registration which involved initial template mesh deformation using Laplacian Surface Deformation (LSD) method guided by a set of uniquely identifiable landmarks. The second step was local shape matching which involved final mesh adaptation using Self-Organizing Maps (SOM) method. Compared to other published methods the landmark correspondence obtained in our method was better (1.1 mm, SD=0.5) with acceptable values of compactness (11 modes), generalization ability (1.6-1.8 mm) and specificity (1.5-2.1 mm).