A Novel Automated Left Ventricle Segmentation Routine

N. Liu, A. Trakic, F. Liu, B. Appleton, S. Wilson, R. Slaughter, W. Strugnell, S. Crozier School of Information Technology and Electrical Engineering, The University of Queensland, Brisbane, QLD, Australia, Cardiovascular MRI Research Centre, The Prince Charles Hospital, Brisbane, QLD, Australia Synopsis In this paper we present a system for the automatic segmentation of the left ventricle (LV) of the heart from breath hold MRI images with subsequent ejection fraction (EF) calculations. The ventricular luminal contours in short and long-axis slices are enhanced with morphological image processing methods and extracted using a novel Global Optimal Closed Path algorithm. In this study, contours both including and excluding LV trabeculations and papillary muscles are considered. Ventricular 3D-reconstructions are based on the use of both short-axis and long-axis contours, with the long-axis contours also being used as an internal skeleton template of the LV to correct for through plane motion. A series of volumes of the reconstructed 3D ventricle is evaluated at different times during the cycle and the EF calculated. By comparing our numerical results with those derived from manual segmentations on eight normal subjects, we conclude that the automated system performance is reliable and consistent. Introduction The measurement of stroke volumes and EF’s from MR images is usually made using manual or semi-automated techniques. The determination of blood volume through reliable computer automated methods is an important and challenging task [1]. Our aim was therefore to design and optimize an automated system to precisely segment the LV and compute the EF. The project included two parts: contour extraction and 3D shape reconstruction from datasets acquired during breath hold on a GE 1.5T Signa Twinspeed Excite platform. These datasets consisted of cine MR image sets obtained using a steady-state free precession imaging sequence (FIESTA) with 20 cardiac phases per slice location. This resulted in 180 short-axis images (9 slice locations) and 20 long-axis images (1 slice location). Contour Extraction After the region of interest was cropped, image enhancement methods such as contrast adjustment, image blurring and others were used to enhance the target (lumen). Morphological operations to remove noise and refine the slices were then applied. For the long-axis slices, a moment theory method was used to register the images. The contour of the endocardium in the short-axis slice is known to be closed and close to round in nature. However, this is not always the case in some cardiac MR images, as the contour is distorted at some locations and can be difficult to determine when conventional gradient methods are used. In order to reconstruct the missing contour around the distorted contour region a variant of the Circular Shortest Path Algorithm was implemented [2, 3]. This algorithm is very successful in searching for the shortest paths of close to circular morphology in complex image terrain. If the contour is continuous and relatively strong, the path will go along it. However in regions where there is no contour or the contour is very weak, the algorithm relies on circular curves to represent the contour, and finally it evaluates the globally optimal path. After pre-processing the image, we computed the gradient of the image and used a corresponding energy functional to compose a cost distribution image. We then transferred the Cartesian to Polar coordinates and mapped the image to the grids. Finally, we applied dynamic programming to search for the global minimum cost (closed curve) path, as the starting and ending nodes are the same points. The above method is very robust for extracting contours that are roughly circular in shape, but does not work well on the long-axis slices where the contours are irregular and arbitrary in shape.