Proceedings Article | 5 May 2004
Proc. SPIE. 5367, Medical Imaging 2004: Visualization, Image-Guided Procedures, and Display
KEYWORDS: Liver, Detection and tracking algorithms, Tumors, Visualization, Surgery, Image registration, Tomography, Data acquisition, Rigid registration, Natural surfaces
To compensate for soft-tissue deformation during image-guided
surgical procedures, non-rigid methods are often used as
compensation. However, most of these algorithms first implement a
rigid registration to provide an initial alignment. In liver tumor
resections, the organ is deformed on a large scale, causing visual
shape change on the organ. Unlike neurosurgery, there is no rigid
reference available, so the initial rigid alignment is based on
the organ surface. Any deformation present might lead to
misalignment of non-deformed areas. This study attempts to
develop a technique for the identification of organ deformation
and its separation from the problem of rigid alignment. The basic
premise is to identify areas of the surface that are minimally
deformed and use only these regions for a rigid registration. To
that end, two methods were developed. First, the observation is
made that deformations of this scale cause noticeable changes in
measurements based on differential geometry, such as surface
normals and curvature. Since these values are sensitive to noise,
smooth surfaces were tesselated from point cloud representations.
The second approach was to develop a cost function which rewarded
large regions with low closest point distances. Experiments were
performed using analytic and phantom data, acquiring surface data
both before and after deformation. Multiple registration trials
were performed by randomly perturbing the post-deformed surface
from a ground truth position. After registration, subsurface
target positions were compared with those of the ground truth.
While the curvature-based algorithm was successful with analytic
data, it could not identify enough significant changes in the
surface to be useful for phantom data. The minimal distance
algorithm proved much more effective in separating the
registration, providing significantly improved error measurements
for subsurface targets throughout the whole surface.
