Cardiac Image Segmentation using Level Sets with Preserved Topology

Cristobal Arrieta^1,2, Sergio Uribe^2,3, Vicente Parot^1,2, Pablo Irarrazaval^1,2, Carlos Sing-Long², Cristian Tejos^1,2

Cardiac performance is typically evaluated from left ventricle volume estimations obtained by segmenting cardiac MRI data. This segmentation is usually done by a tedious manual process, that shows low reproducibility. Different authors have proposed the use of Level-Sets to automate this process. Due to the presence of multiple objects in the image, Level-Sets normally require shape constrains that are built from a training procedures. Training-based segmentations tend to fail with severely abnormal anatomies. We propose the use of Level-Sets with Preserved Topology. Compared to manual segmentations we achieved equally accurate and more reproducible segmentations without training, even in abnormal patients.