Weighted Labels for 3D Image Segmentation

Abstract

Segmentation tools in medical imaging are either based on editing geometric curves or on the assignment of region labels to image voxels. While the first approach is well suited to describe smooth contours at subvoxel accuracy, the second approach is conceptually more simple and guarantees a unique classification of image areas. However, contours extracted from labeled images typically exhibit strong staircase artifacts and are not well suited to represent smooth tissue boundaries. In this paper we describe how this drawback can be circumvented by supplementing region labels with additional weights. We integrated our approach into an interactive segmentation system providing a well-defined set of manual and semi-automatic editing tools. All tools update both region labels as well as the corresponding weights simultaneously, thus allowing one to define segmentation results at high resolution. We applied our techniques to generate 3D polygonal models of anatomical structures.