Abstract
The integration of statistical shape models (SSMs) for generating a patient-specific model from sparse data is widely spread. The SSM needs to be initially registered to the coordinate-system in which the data is acquired and then be instantiated based on the point data using some regressing techniques such as principal component analysis (PCR). Besides PCR, partial least squares regression (PLSR) could also be used to predict a patient-specific model. PLSR combines properties of PCR and multiple linear regression and could be used for shape prediction based on morphological parameters.
Both methods were compared on the basis of two SSMs, each of them constructed from 30 surface models of the proximal femur and the pelvis, respectively. Thirty leave-one-out trials were performed, in which one surface was consecutively left out and further used as ground truth surface model. Landmark data were randomly derived from the surface models and used together with the remaining 29 surface models to predict the left-out surface model based on PCR and PLSR, respectively. The prediction accuracy was analysed by comparing the ground truth model with the corresponding predicted model and expressed in terms of mean surface distance error.
According to their obtained minimum error, PCR (1.62 mm) and PLSR (1. 63 mm) gave similar results for a set of 50 randomly chosen landmarks. However PLSR seems to be more susceptible to a wrong selection of number of latent vectors, as it has a more variation in the error.
Although both regression methods gave similar results, decision needs to be done, how to select the optimal number of regressors, which is a delicate task. In order to predict a surface model based on morphological parameters using PLSR, the choice of the parameters and their optimal number needs to be carefully selected.
