Regresar

Automatic segmentation of the left atrium in cardiac computed tomography

Abstract:

The present work proposes a technique for the automatic segmentation of the left atrium (LA) in 10 three-dimensional (3-D) cardiac images of multi-cut computed tomography, belonging to the same subject. The mentioned technique consists of the stages of pre-processing and segmentation. The pre-processing step includes two phases. In the first phase, in order to minimize both Poisson noise and the impact of the staircase artifact, a technique called global similarity enhancement is used. This type of enhancement consists in the application of a bank of filters, softeners and a border detector, whose purpose is to generate an image in which the information of the anatomical structures that make up the original images are grouped. In the second phase, considering the filtered images, we use a priori information about the location of the mitral valve and a learning paradigm, based on vector support machines, to define a region of interest that isolates LA from neighboring anatomical structures. On the other hand, to generate the 3-D morphology of the left atrium, a segmentation stage is applied which considers the pre-processed images and a clustering algorithm based on regions growth. The proposed strategy generates 3-D segments of the left atrium in all images that make up the complete cardiac cycle of the subject considered. In order to quantify the performance of the referred technique, the Dice coefficient was considered, obtaining a good correlation between the automatic segmentations and the manual ones generated by a cardiologist.