An implementation of independent component analysis for 3D statistical shape analysis

•A unique ICA formulation for complex multi-dimensional continuous signals is shown.•A novel framework to incorporate ICA for statistical shape analysis is shown.•The ICA algorithm is validated for decomposition of 3D surfaces.•ICA is compared to PCA for human right ventricle decomposition.

An implementation of the independent component analysis (ICA) technique for three-dimensional (3D) statistical shape analysis is presented. The capabilities of the ICA approach to uncover inherent shape features are first demonstrated through analysis of sets of artificially generated surfaces, and the nature of these features is compared to a more traditional proper orthogonal decomposition (POD) technique. For the surfaces generated, the ICA approach is shown to consistently extract surface features that closely resembled the original basis surfaces used to generate the artificial dataset, while the POD approach produces features that clearly mix the original basis. The details of an implementation of the ICA approach within a statistical shape analysis framework are then presented. Results are shown for the ICA decomposition of a collection of clinically obtained human right ventricle endocardial surfaces (RVES) segmented from cardiac computed tomography imaging, and these results are again compared with an analogous statistical shape analysis framework utilizing POD in lieu of ICA. The ICA approach is shown to produce shape features for the RVES that capture more localized variations in the shape across the set compared to the POD approach, and overall, the ICA approach produces features that represent the RVES variation throughout the set in a considerably different manner than the more traditional POD approach, providing a potentially useful alternate to statistically analyze such a set of shapes.