Left Ventricle Segmentation In Cardiac MRI Based On Improved Sparse Shape Composition Model

Cardiac function assessment based on left ventricle(LV) contour in cardiac MR images, is an important measurement in the diagnosis and treatment monitoring of cardiovascular diseases. Thus, LV endo- and epicardium extraction has become a hot topic in this field of study. Affected by the noise and misleading information, methods solely relying on image appearance cues usually tend to achieve unsatisfactory result. Recent years, shape priori knowledge has proven to be helpful to improve the segmentation accuracy and robustness in many studies. Shape prior modeling plays a significant role in these processes.Most existing methods focuses on single shape modeling. When modeling LV endo- and epicardium separately, the spatial information between them is effectively utilized. Thus for those input shape with large or relatively dense errors, these methods are difficult to achieve better results. In this paper, an extension of Sparse Shape Composition(SSC) model, in which multiple interested shapes are regarded as a group and modeled simultaneously is proposed. For an input group, an optimal sparse linear combination of existing training groups is calculated to approximate it. Thus, not only the a-priori information of each shape but also the a-priori correlativity information among them is implicitly incorporated on-the-fly.Compared with the original SSC, an initial epicardium can be estimated by this method without complicated landmark detection, which has proven to be valuable in the deformation process. On this basis, a-prior based epicardium segmentation is implemented. First, a reasonable epicardial contour is inferred based on improved SSC. With such estimation as constraint shape, deformation based on the adopted shape-constraint GCV model is conducted to drive the evolution contour converging the proper epicardium position. After that, an iterative process of refinement and deformation is conducted to acquire the epicardium segmentation. While refinement based on improved SSC is aiming at regularizing the noisy deformed contour, whose result is utilized as the initial position and constraint shape in the next GCV based deformation.Experiment results show that it can be used to localize and extract LV epicardium effectively. Localization error is around 2.74 m. Segmentation error is about 1.50 mm.