FAST INTERACTIVE MULTI-REGION CARDIAC SEGMENTATION WITH LINEARLY ORDERED LABELS

We present a novel and fast interactive approach to multi-modality cardiac image segmentation, which employs the linearly ordered surfaces as an additional constraint. We show using such a geometrical constraint helps to significantly reduce user interaction and improve the accuracy of segmentation results at the same time. We solve the proposed multi-region segmentation problem with the order constraints by means of convex optimization, resulting in a fast and reliable flow maximization approach which implicitly embeds the linear order prior without introducing extra computation load. In this regard, a new fully parallelized continuous max-flow algorithm is proposed and implemented using GPGPU to segment a 3D volume within one second. We demonstrate our results over pathological trans-esophageal echocardiogram, cardiac CT and delayed enhancement MRI data sets.