bul 2004
sfined as
/ the side
inductive
ixel, the
line are
line are
growing
er-drawn
re area is
vhere the
L several
he results
Is for the
de of this
nple, the
result of
ight-grey
rawn line
B and C
e line are
y with 12
students
, six with
itaset was
0 create a
ve region
ower than
ing, local
Dataset 4,
where the
This is
se the old
nplicated,
1
o : : | docu
1 2 3 4 5
Figure 6. Average times in minutes needed for a complete
segmentation of five different CT datasets. The blue
line with diagonal squares shows the results for the
old version of the tools, the other for the new
version.
To compare the reproducibility of the segmentation results, we
calculated the intra-observer accuracies by means of the
Tanimoto coefficient (Tanimoto, 1958). This coefficient is 100
percent if two segmentations are completely identical and 0 if
they do not overlap at all. As figure 7 shows, the intra-observer
accuracies are on average maintained at the same high level as
with the reference tools.
98,0%
97,5%
e : ~/ : |
95,596
95,0%
94,5% TRE : :
1 2 3 4 5
Figure 7. Intra-observer variabilites for five different
datasets, measured with the Tanimoto coefficient.
The blue line with diagonal squares shows the
results for the old version of the tools, the other for
the new version.
Additionally to these quantitative results for time and accuracy,
we logged the usage of the new region grower tool to reveal
more information regarding the preferred modes the users have
been working with. Figure 8 shows the results of this analysis,
where each mouse click with the region grower was counted as
one event. As can be seen, the users took advantage of the Fix
button to work with fixed borders of the grey value interval
quite frequently. According to the diagram, only about half of
all correction attempts were successful. In reality, the number
where a leaked region cannot be removed is much lower, but
users tended to try the removal repeatedly and clicked several
times in these cases.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004
Use of the new region grower
25 O Interactive borders
15% ( gi Fixed borders
Oo Failed corrections
; B Succesful corrections
Figure 8. Use of the new region growing tool.
4. CONCLUSIONS
We introduce a new, interactive approach to detect and correct
leaks in region growing segmentation that is suitable for two-
and three-dimensional datasets alike. Moreover, we present a
novel tool to manually correct the contours of segmented
objects which minimizes the necessary user interaction.
The new user interface employed in these tools, featuring a
direct feed-back of parameter changes, is approved by the users
and leads to a higher acceptance of the region growing tool.
The times needed for a liver segmentation in the clinical
workflow could be reduced noticeable using the methods
described in this paper.
REFERENCES
Arya, S., D.M. Mount, N.S. Netanyahu, R. Silverman and A.Y.
Wu, 1998. An optimal algorithm for approximate nearest
neighbor searching fixed dimensions. Journal of the ACM,
45(6), pp. 891-923.
Dalheimer, M.K., 2002. Programming with Qt. O'Reilly, 2.
edition.
Engelmann, U., A. Schröter, M. Schwab, U. Eisenmann, L.
Bahner, S. Delorme, H. Hahne and H.-P. Meinzer, 2000. The
Linux-based PACS project at the German Cancer Research
Center. In: Proc. CARS 2000, Elsevier, Amsterdam, pp. 419-
424.
Gonzalez, R.C. and R.E. Woods, 2002. Digital Image
Processing. Addison Wesley, 2"? edition.
Heckbert, P., 1990. Concave Polygon Scan Conversion. In:
Glassner, A. (editor): Graphics Gems. Academic Press, Boston,
pp- 87-91.
Kunert, T., T. Heimann, A. Schróter, M. Schóbinger, T.
Bóttger, M. Thorn, I. Wolf and H.-P. Meinzer, 2004. An
Interactive System for Volume Segmentation in Computer-
Assisted Surgery. In: Proc SPIE Medical Imaging 2004.
