Subject Specific Shape Modeling with Incremental Mixture Models

摘要

Statistical shape models provide versatile tools for incorporating statistical priors for image segmentation. Difficulties arise, however, when the target anatomical shape differs significantly from the training set used for model construction. This paper presents a novel approach for fast and accurate segmentation of subject-specific geometries based on models largely derived from normal subjects. This technique is particularly suitable for analyzing complex structures such as severely abnormal patient datasets. The proposed method uses online principal component update to incorporate subject-specific geometry. Mixture models are used to estimate the latent density distribution of the data, thus enabling adequate constraining during active shape propagation. Validation based on hypertrophic cardiomyopathy (HCM) datasets with MRI shows significant improvement in overall accuracy and increased adaptation to complex structures.