文摘
Statistical shape modeling (SSM) is a widely popular framework in cardiac image analysis, especially for image segmentation and computer-aided diagnosis. However, the conventional PCA-based models produce new axes of variation which are statistically motivated but thus are not necessarily clinically meaningful. In this paper, we propose an alternative method for statistical decomposition of the shape variability based on partial least squares (PLS). With this method, the model construction is achieved such that it is constrained by the specific clinical question of interest (e.g., estimation of disease state). To achieve this, instead of deriving modes of variation in the directions of maximal variation as in PCA, PLS searches for new axes of variation that correlate most with some output clinical response variables such as diagnostic labels, leading to a decomposition that is anatomically and clinically more meaningful. The validation carried out with 200 cases from the Cardiac Atlas Project database as part of the MICCAI 2015 challenge on SSM, including healthy and infarcted left ventricles, shows the strength of the proposed PLS-based statistical shape model, with 98 % prediction accuracy.