Nonlinear filtering for extracting orientation and tracing tubular structures in 2-D medical images

We consider the problems of extracting local fiber orientation and tracing tubular structures in 2-D medical images. We first present a nonlinear filter for detecting overlaid orientations at each pixel. The filter is used to build two profiles, the directional profile and the appearance profile, which are designed to give high responses along an oriented structure. The local orientation at a pixel is then obtained from the local maxima of the product of these two profiles. We also develop an algorithm for tracing tubular structures. Starting from a user-specified seed point on the fiber, we find the next point on the fiber by locating the maxima of the tracing profile, which favors the smoothness of the curve as well as its alignment with the estimated orientation. We evaluate the proposed method on different medical structures such as cardiac myofibers, vessels, and microtubules.