3D/4D cardiac segmentation using active appearance models, non-rigid registration, and the Insight Toolkit

We describe the design of a statistical atlas-based 3D/4D cardiac segmentation system using a combination of active appearance models (AAM) and statistical deformation models with the Insight Toolkit as an underlying implementation framework. Since the original AAM approach was developed for 2D applications and makes use of manually set landmarks its extension to higher dimensional data sets cannot be easily achieved. We therefore apply the idea of statistical deformation models to AAMs and use a deformable registration step for establishing point-to-point correspondences. An evaluation of the implemented system was performed by segmenting the left ventricle cavity, myocardium and right ventricle of ten cardiac MRI and ten CT datasets. The comparison of automatic and manual segmentations showed encouraging results with a mean segmentation error of 2.2±1.1 mm. We conclude that the combination of a non-rigid registration step with the statistical analysis concepts of the AAM is both feasible and useful and allows for its application to 3D and 4D data.