ibul 2004
eview of
utline of
erator to
loreover,
ly texture
the real-
tation is
nentation
ating the
lgorithms
algorithm
tual pixel
are four
following
elow the
there are
situation
ark pixel,
'cognized
extraction
s. If there
(A), the
becomes
he object
lored as a
ht. If this
"it 1s part
the white
ght. Else,
does not
action on
luded the
n, using a
>xtraction.
ion of the
jeted are
tracted is
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004
2.1.3 Leak detection: Even when using an optimal grey
value interval, the region growing algorithm can leak out if
contrast at the object boundary is not sufficient. In case of the
liver, especially the regions next to stomach and kidney are
susceptible to leaking, since these organs have similar grey
values as the liver itself. Often, the origin of the leak is only a
narrow connection between neighbouring tissues. Since the
time region growing algorithms are employed in segmentation,
solutions that are both fast and easy to control have been
searched to solve this problem. Often, the segmentation
software allows the user to draw a line which blocks the region
growing and thus prevents leakage. However, the most user-
friendly approach would require just one mouse click in the
leak area to remove it. The only proposed solution with this
interaction scheme (Sekiguchi and Sano, 1994) has a major
draw-back: it only works with very narrow connections.
We propose a new method using distance transformation to
identify leak regions and remove the additional area. The basic
idea is as follows: From an arbitrary point within the erroneous
area, we calculate the path to the seed point which always
maintains the maximum possible distance to the surrounding
edges. The point on this path which after a local maximum has
the shortest distance to the contour marks the narrowest
bottleneck to the chosen region and thus the origin of the leak.
The constraint that a local maximum has to be passed first
prevents that the starting point of the path is marked as origin if
the user clicked at a position too close near the edge.
A direct implementation of this idea uses the skeletonization or
medial axis transformation (Lee, 1982) of the object of interest.
There are several different definitions for the MAT of a shape,
one being the following: assuming the shape is of a burnable
material and is set on fire simultaneously on all sides, it will
burn down towards the center. Where several fronts of fire
meet, they extinguish themselves and the resulting line graph is
the skeleton of the shape.
Figure 2. Skeletonization of a leaked liver shape by
thresholding the gradient magnitudes of the distance
image. All pixels with a gradient magnitude lower
than 0.95 are marked black.
When the path follows this skeleton as much as possible, the
maximum distance to the object border is maintained
automatically. A fast implementation to test if points belong to
the skeleton or not is based on the gradient magnitudes of the
distance image of the shape, which can efficiently be calculated
using the algorithm presented in (Arya, 1998). All points where
this magnitude is zero belong to the skeleton. Due to
inaccuracies in the distance transformation of quantized images,
it is better in practice to test if the gradient magnitude lies
below a specific threshold. For an example of this method see
figure 2.
Being able to determine if a pixel is part of the skeleton or not,
the next step consists of finding the shortest path over this
skeleton. An extensive graph search on the entire skeleton is the
most obvious solution to this problem, but to save the inherent
costs of computation, we opted for a different method. Our
algorithm makes use of the fact that the path always leads from
the periphery towards the seed-point. During the region
growing phase we count at which step each pixel is added to the
region, defining a geodetic distance function. During the path
search, we only consider pixels with a geodetic distance lower
or equal to the distance at the current position and chose among
those the one with the lowest gradient magnitude. In case the
lowest magnitude is lower than 0.5 (i.e. the path has not
reached the skeleton yet), we alternatively chose the pixel with
the largest distance to the contour.
For every pixel in the path, we check its distance to the contour
for a local minimum to find bottlenecks of the shape. The
smallest local minimum is marked as the origin of the leak.
Subsequently, we search the two nearest, opposing points on
the contour at the acquired position and separate the segmented
area along a line between these two points. The part of the area
where the user clicked is then removed from the segmentation.
An example of this operation is shown in figure 3.
Figure 3. The region grower with seed point in the liver
(white circle) has leaked out into the stomach. The
path calculated by the heuristic from the point of
correction is displayed yellow-green. The blue line
shows where the contour will be separated at the
leak origin.
2.2 The new contour correction tool
The presented leak detection for the region growing tool works
in cases with a bottleneck between leaked area and origin. If
this bottleneck is not present or the deviation is too small the
method will not work. Additionally, as we found out in the user
survey of our segmentation software, demand for an efficient
correction tool is very high. The basic freehand segmentation,
i.c. adding or subtracting an enclosed area, is not sufficient:
Changing between the two modes is time-consuming and error-
prone, moreover, it is cumbersome to revolve the entire
correction area.