ICA vs. PCA Active Appearance Models: Application to Cardiac MR Segmentation

Abstract

Statistical shape models generally use Principal Component Analysis (PCA) to describe the main directions of shape variation in a training set of example shapes. However, PCA has the restriction that the input data must be drawn from a Gaussian distribution and is only able to describe global shape variations. In this paper we evaluate the use of an alternative shape decomposition, Independent Component Analysis (ICA), for two reasons. ICA does not require a Gaussian distribution of the input data and is able to describe localized shape variations. With ICA however, the resulting vectors are not ordered, therefore a method for ordering the Independent Components is presented in this paper. To evaluate ICA-based Active Appearance Models (AAMs), 10 leave-15-out models were trained on a set of 150 short-axis cardiac MR Images with PCA-based as well as ICA-based AAMs. The median values for the average and maximal point-to-point distances between the expert drawn and automatically segmented contours for the PCA-based AAM were 2.95 and 8.39 pixels. For the ICA-based AAM these distances were 1.86 and 5.01 pixels respectively. From this, we conclude that the use of ICA results in a substantial improvement in border localization accuracy over a PCA-based model.

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.