Clustering based 3D level set method for volumetric cardiac segmentation

Abstract

Multi-slice CT (MSCT) provides dynamic three-dimensional (3D) volumetric data of the whole heart, and is an important medical imaging tool for diagnosis of cardiac diseases. Due to the large size of the dynamic data, manual identification, segmentation and tracking of various parts of the heart will be very labor intensive and inefficient. Alternatively, sophisticated image processing techniques, which require minimal user intervention, can be developed and employed to automate such tasks. In this work, we propose a semi-automatic clustering based 3D level set method to robustly segment the endocardium surface from cardiac MSCT images. The theory of level set defines a flexible and powerful surface which is capable of capturing the complex endocardium anatomical structure. A novel speed function for the level set method using a clustering algorithm is proposed to exploit the non-homogeneous blood pool intensity property by supporting a set of independent intensity samples. To define the intensity clusters for the blood pool region and the surrounding region, only a few lines drawn on the corresponding regions are required as user input. The segmentation result is a level set 3D surface in the whole volume space which can readily be constructed to form a spatial model. Our clustering based 3D level set method can also be used for segmenting other heart wall surfaces by performing appropriate initialization. By ex-tending to a 4D level set method, 4D (3D plus time) dynamic volumetric data could be readily processed