methods that require some level of user input. These methods provide more accuracy
and more control over automatic methods. Work in this area involves using an input
device (i.e. the mouse) and marking the segments manually. Given enough user
input, automatic techniques can then be applied to the volume with greater accuracy.
The additional input from the user helps guide the automatic process, and using
automation prevents the user from manually marking the entire volume, which can
be an error-prone and time consuming process.
Previous works on semi-automatic methods include Owada et ah’s Volume Catcher
and Yuan et al.’s Volume Cutout [19, 29]. Volume Catcher allows the user to draw
2D free-form strokes on volume; the strokes are then extended using region growing
and set as constraints for graph-based segmentation. Volume Cutout uses two kinds
of strokes to denote the foreground and the background on a two-dimensional view
of the 3D volume. The strokes are then used in a graph-based approach to automate
the segmentation for the rest of the volume. These two approaches both focus on
the problem of two-material segmentation. We are interested in generalized, multi-
material segmentation.
Another difficulty in semi-automatic segmentation is in manipulating 3D objects
using a 2D interface. Many of the existing methods involve extending 2D segmen-
tation techniques by applying segmentation to 2D slices and re-constructing the 3D
segmentation. However, it is hard to identify 3D spatial correlation and features using
this approach. We choose an approach that operates directly on the 3D volume with
semi-automated segmentation using graph-cut. Graph-cut methods have proven to
be useful for segmenting 2D images. Typically, 2D graph-cuts involve denoting the
foreground and background of the graph∕image through user input and performing
a min-cut/max-flow variant to identify two sets of nodes∕pixels. The min-cut in-
duces a natural component partition across the image that satisfies the criterion of
segmentation. The 3D complement of this approach has been used by Liu et al. for
bone segmentation [14]. Their work takes the rectilinear grid as the graph for the