Reusability of Statistical Shape Models for the Segmentation of Severely Abnormal Hearts

Statistical shape models have been widely employed in cardiac image segmentation. In practice, however, the construction of the models is faced with several challenges, in particular the need for a sufficiently large training database and a detailed delineation of the training images. Moreover, for pathologies that induce severe shape remodeling such as for pulmonary hypertension (PH), a statistical model is rarely capable of encoding the significant and complex variability of the class. This work presents a new approach for the segmentation of abnormal hearts by reusing statistical shape models built from normal population. To this end, a normalization of the pathological image data is first performed towards the space of the normal shape model, which is then used to guide the segmentation process. Subsequently, the model recovered in the space of normal anatomies is propagated back to the pathological images space. Detailed validation with PH image data shows that the method is both accurate and consistent in its segmentation of highly remodeled hearts.