Interface detection in diffusion tensor MRI

Abstract

We present a new method for detecting the interface, or edge, structure present in diffusion MRI. Interface detection is an important first step for applications including segmentation and registration. Additionally, due to the higher dimensionality of tensor data, humans are visually unable to detect edges as easily as in scalar data, so edge detection has potential applications in diffusion tensor visualization. Our method employs the computer vision techniques of local structure filtering and normalized convolution. We detect the edges in the tensor field by calculating a generalized local structure tensor, based on the sum of the outer products of the gradients of the tensor components. The local structure tensor provides a rotationally invariant description of edge orientation, and its shape after local averaging describes the type of edge. We demonstrate the ability to detect not only edges caused by differences in tensor magnitude, but also edges between regions of different tensor shape. We demonstrate the method's performance on synthetic data, on major fiber tract boundaries, and in one gray matter region. © Springer-Verlag Berlin Heidelberg 2004.