No abstract

We present a robust, fast and fully automatic approach enabling the segmentation of the main anatomical structures of the heart in CT images. The proposed method is based on the adaptation of a 3D triangulated mesh to new unknown images exploiting simultaneously knowledge of organ shape and typical gray level appearance in images, both learned from a training database made of 28 data sets. The described approach was tested on more than 50 volume images at different cardiac phases. Visual inspection by experts reveals that the proposed method is overall robust and succeeds in segmenting the heart up to minor interactive local corrections. IntroductionAutomatic segmentation plays a central role when inspecting reconstructed 3D cardiac images (e.g. from CT or MR scanners) [1]. An accurate classification of the different cardiac regions is usually the first step of subsequent tasks like: Visualization, coronary artery inspection, measurement of the ejection fraction for the left and right ventricles, wall motion analysis, intervention planning (e.g. for electro-physiology treatment).Because a significant amount of information about global functional analysis is delivered by the left ventricle, much work has been dedicated in extracting this structure from medical images. However, volumetric images acquired by emerging imaging techniques like e.g. multislice CT offer much more information, motivating the need for methods that are able to segment the other anatomical structures of the cardiac region.In this paper, we present a model-based approach capable of extracting the main anatomical structures of the heart. Unlike other works on model-based segmentation, we also concentrated our efforts in procedures enabling the initialization of the model without any click required from the user.During the training phase, typical boundary appearances are learned from representative images. This allows a very robust boundary detection when segmenting new unknown images. Moreover, although we adapt a single mesh, special attention has been paid in allowing each anatomical part of the model to be free to globally deform in an independent fashion. 2.Shape-constrained deformable models frameworkDeformable models have been widely used for the segmentation of medical images [2]. However, they might have too much flexibility when adapting towards the organ of interest and be sensitive to image artifacts. Using a priori information about shape variability to constrain the deformation flexibility has been recognized to improve the robustness of the segmentation process [3]. In this paper, we will use the approach introduced in [4], which constrains the model to remain close to the trained shape while allowing local non-learned deformations to account for the individuality of each patient.The input of the shape-constrained deformable model approach is a mesh with vertices v j (j = 0...V ) connected in T triangles. The mesh is adapted to a new image minimizing an energy function, which is usually made of two contributions. The fir...

No abstract