Supervised Segmentation of the Circle of Willis in 7-Tesla MRI using Variance Wells

Abstract

This paper describes a novel process for segmenting cerebral vasculature in 3D images based on a construct we call variance wells (or vWells), which are local clusters of relatively homogeneous voxels that preserve anatomical boundaries. We present a semiautomated algorithm to combine adjacent vWells to efficiently segment blood vessels, in an interactive 3D environment guided and visually validated by a human operator. This provides an alternative to labor-intensive methods involving manual or semiautomated tracing of 2D cross sections to obtain accurate vessel segmentation. The automated portion of our segmentation process builds on well-established statistical methods, namely, the Welch’s t-test, to group neighboring vWells belonging to a given anatomical object, and to distinguish them from the vWells in the background immediately surrounding the object. The manual portion of the process employs a semitransparent surface rendering of the segmentation superimposed on a movable slice of the image data, enabling efficient guidance of the process and precise supervision of the result. We demonstrate the method by segmenting the Circle of Willis in 27 time-of-flight MRI images from a 7-Tesla scanner.