DTMRI Segmentation using DT-Snakes and DT-Livewire

Abstract

In this paper we extend two popular classical scalar medical image segmentation techniques to diffusion tensor magnetic resonance images (DTMRI). We propose DT-snakes and DT-livewire through modifying the external image forces in snakes and cost terms in livewire. The new forces and cost terms are derived from and operate on a DT field rather than a scalar image. This is achieved by making use of recent advances in DT calculus and DT dissimilarity measures, as well as DT smoothing and DT interpolation. Proper quantification of tensor dissimilarity allows for defining spatial gradient vectors and gradient magnitudes of DT fields, an essential component for attracting snakes or livewire to target boundaries in DT images. DT calculus enables weighted averaging of tensors which is essential for both pre-smoothing of DT images prior to segmentation, as well as interpolation of tensors on non-grid positions in the image. We evaluate different recent DT tensor dissimilarity metrics including the Log-Euclidean and the square root of the J-divergence. We present qualitative and quantitative DT segmentation results on both synthetic and real cardiac and brain DTMRI data