A Substructure Segmentation Method of Left Heart Regions from Cardiac CT Images Using Local Mesh Descriptors, Context and Spatial Location Information

Abstract

Most of the traditional segmentation algorithms for cardiac medical images are semi-automatic. The procedures are subjective, hardly reproducible and time-consuming. With rapid growth of medical images, fully automatic and reliable segmentation methods are desirable for the quantitative and massive analysis. In this paper, we propose a novel approach for automatic segmentation of left heart substructures from cardiac CT images. The approach consists of two major processes. The first process is the threshold-based image segmentation which extracts left heart regions from cardiac CT images and the 3D mesh data are extracted from the image segmentation results consequently. The second process is a 3D mesh segmentation which further partitions the mesh data into the substructures of the left heart. In the process of 3D mesh segmentation, the heart mesh model is first processed by ant colony optimization algorithm to generate small patches. Further, this paper uses a supervised learning method to further segment the left heart regions into substructures. In addition to the local features of the patches, the input vectors for the supervised learning also contain the context features of the patches and the spatial location features of the patches. The actual CT images are used to validate the proposed method, the results are compared with other methods such as the k-means clustering and normalized cut. It turns out that the proposed method can obtain improved results. The combination of image segmentation and 3D mesh segmentation is a novel methodology in segmenting complex cardiac substructures. When using a supervised learning, the use of context features and the spatial location features can help to distinguish some substructures while the use of local features are not sufficient.