Automatic Segmentation of White Matter Structures from DTI Using Tensor Invariants and Tensor Orientation

INTRODUCTION DTI analysis of brain structures has shown to be relevant in a number of neurological clinical pathologies, such as brain ischemia, multiple sclerosis or epilepsy, among others. In schizophrenia, group studies have demonstrated alterations in the diffusion of several fiber bundles within the white matter [1]. The automatic segmentation of these structures from DTI has spurred significant research effort recently, due to its importance for these studies. However, automatic segmentation from DTI is a very challenging issue, and approaches in the literature require the manual identification of a region of interest or starting seeds for the segmentation to be performed. This abstract presents a fully automatic segmentation approach for different anatomical structures in the white matter, based on: (a) the use of tensor invariants, together with the orientation of the tensor, as features to drive the segmentation, (b) a level set formulation based on the statistical modeling of the data though vector-valued Geodesic Active Regions (GAR), and (c) the use of a MRI atlas of human white matter [2], which is nonlinearly registered to the volume under study, to automatically obtain initial contours for the segmentation process. Segmentation on several DTI volumes showed good results, demonstrating that a fully automatic segmentation is possible and also presenting good properties when compared to recent DTI segmentation approaches in the literature. METHODS The proposed method is based on a vector-valued version of the GAR level set segmentation method [3]. Let v(x) be the feature vector at voxel x. Then, we seek to minimize the following energy functional: