Automatic Heart Segmentation using Gradient-assisted Localized Active Contour in Cardiac CT Images

Abstract

The heart is one of the most important human organs, and composed of complex structures. Computed tomography angiography (CTA), magnetic resonance imaging (MRI), and single photon emission computed tomography are widely used, non-invasive cardiac imaging modalities. Compared with other modalities, CTA can provide more detailed anatomic information of the heart chambers, vessels, and coronary arteries due to its higher spatial resolution. To obtain important morphological information of the heart, whole heart segmentation is necessary and it can be used for clinical diagnosis. In this paper, we propose a novel framework to segment the four chambers of the heart automatically. First, the whole heart is coarsely extracted. This is separated into the left and right parts using a geometric analysis based on anatomical information and a subsequent power watershed. Then, the proposed gradient-assisted localized active contour model (GLACM) refines the left and right sides of the heart segmentation accurately. Finally, the left and right sides of the heart are separated into atrium and ventricle by minimizing the proposed split energy function that determines the boundary between the atrium and ventricle based on the shape and intensity of the heart. The main challenge of heart segmentation is to extract four chambers from cardiac CTA which has weak edges or separators. To enhance the accuracy of the heart segmentation, we use region-based information and edge-based information for the robustness of the accuracy in heterogeneous region. Model-based method, which requires a number of training data and proper template model, has been widely used for heat segmentation. It is difficult to model those data, since training data should describe precise heart regions and the number of data should be high in order to produce more accurate segmentation results. Besides, the training data are required to be represented with remarkable features, which are generated by manual setting, and these features must have correspondence for each other. However in our proposed methods, the training data and template model is not necessary. Instead, we use edge, intensity and shape information from cardiac CTA for each chamber segmentation. The intensity information of CTA can be substituted for the shape information of the template model. In addition, we devised adaptive radius function and Gaussian-pyramid edge map for GLACM in order to utilize the edge information effectively and improve the accuracy of segmentation comparison with original localizing region-based active contour model (LACM). Since the radius of LACM affects the overall segmentation performance, we proposed an energy function for changing radius adaptively whether homogeneous or heterogeneous region. Also we proposed split energy function in order to segment four chambers of the heart in cardiac CT images and detects the valve of atrium and ventricle. In experimental results using twenty clinical datasets, the proposed method identified the four chambers accurately and efficiently. We also demonstrated that this approach can assist the cardiologist for the clinical investigations and functional analysis.