International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004 
2.2.1 Basic Idea: To remedy these problems, the new 
correction tool should work directly on the contours of the 
segmentation. Figure 4 points out the difference to the old 
approach: To delete a leaked region, it should be sufficient to 
draw the cut line. In a similar way, the extension of an existing 
segmentation should be possible by just drawing the new 
contour. Furthermore, it should be possible to combine these 
operations in a single interaction, i.c. by drawing one line, the 
user should be able to cut the region at one place and extend it 
at another. 
  
Figure 4. Required correction lines: cut operation with the 
new (A) and the old approach (B), merge operation 
with the new (C) and the old approach (D). 
2.2.2 Heuristics: In order to implement this idea, we 
developed a set of heuristics to make the necessary decisions 
automatically. The decision if a cut or a merge operation should 
be executed is made on the existing segmentation: If the user 
draws a line in an unsegmented area, the region should be 
extended up to that line, else it should be cut at that line. One 
line sector is defined by two crossings with the old contour. 
Thus, the first steps after drawing a line with the correction tool 
are to find the points of intersection with the contour, split the 
line into the corresponding sectors and classify each sector as 
cut or merge. 
The second heuristic has to decide which part of the region is to 
be modified. Both operations leave two possibilities to interpret 
the new contour. In figure 4a, e.g. it is not clear if the area to 
the left or to the right of the line should be deleted. However, 
since corrections generally are minor adjustments on the current 
result, the new contour is always interpreted in the way that 
results in the smallest area of change. Thus, for a cut operation 
always the smaller part of the region is deleted and for a merge 
operation the smaller part is added. If the two parts do not ditfer 
significantly, no operation is performed. 
2.2.3 Implementation: 
To measure the area on both sides of the drawn line, a modified 
region grower is used which counts all added pixels. It should 
be noted that there are several different possibilities to separate 
the two areas. Using the line drawn by the user can lead to 
problems in the area measurement, since areas may be split in 
several separated regions: In the example shown in figure 5a, 
the area above the line would be split in a five and a two pixel 
region. To avoid the resulting problems, we use two different 
lines to separate the areas, as shown in figures 5b and c. The 
dark-grey colored separators in theses figures are defined as 
direct neighbors of the drawn correction line, sorted by the side 
they belong to. To determine this side, we use an inductive 
procedure: Originating from an arbitrary neighbor pixel, the 
adjacent neighbors are marked until both ends of the line are 
reached. At that point, all pixels on one side of the line are 
known, the rest is assigned to the other side. 
Finally, to determine the size of the area, region growing 
algorithms can be started from every point of the user-drawn 
line and their results are added. In most cases the entire area is 
captured by the first region growing, but in cases where the 
correction line runs close to the border of the object, several 
runs might be necessary. In the example in figure 5, the results 
are an area of 11 pixels for the upper side and 22 pixels for the 
lower (only for the shown section). Thus, the upper side of this 
line sector is filled. For the next sector in the example, the 
lower pixels will be removed and 5d shows the final result of 
the operation. 
A 
  
  
Figure 5. Procedure for a contour correction of the light-grey 
object: The black arrow shows the user-drawn line 
of correction in the original image (A), in B and C 
the areas of the regions on both sides of the line are 
determined, D shows the final result. 
3. RESULTS AND DISCUSSION 
We evaluated the new tools quantitatively in a study with 12 
medical students on five different CT datasets. Six students 
worked with the old version of the segmentation tools, six with 
the new version. In a randomized sequence, every dataset was 
segmented four times by each test person. The times to create a 
complete segmentation of the liver with the interactive region 
grower and the contour modifier are on average 10% lower than 
the ones created with a set of standard region growing, local 
threshold and freehand segmentation (see figure 6). Dataset 4, 
the image with the lowest contrast, is the only one where the 
test users working with the old tools were faster. This is 
because in this case most of them did not bother to use the old 
region grower at all, which they regarded as too complicated, 
and used other tools instead. 
   
Intern 
- 
Figur 
To « 
calcu 
Tanir 
perce 
they 
accur 
with 
98,0% 
97,5% 
97,0% 
96,5% 
96.0% 
95,5% 
95,0% 
94,5% 
Figui 
Addi 
wel 
more 
been 
wher 
one « 
butto 
quite 
all c 
wher 
Users 
times