in spite of 
interactive 
anges the 
the screen 
timal seg- 
old. After 
holes and 
areas, we 
er, 1993). 
ea thresh- 
| might be 
functions 
egmented 
operators 
10les. The 
s selected 
> the con- 
cking this 
automati- 
nd. After 
which are 
ontouring 
on easily. 
vitable to 
literature, 
thod. For 
1 give the 
the inter- 
ce editing 
ntation is 
s days to 
CTION 
Y 
software 
ested the 
ecause of 
or human 
| mask on 
We have 
by using 
1 parame- 
automatic 
. For this 
matching 
areas are 
m, subdi- 
ler image 
olar con- 
mum and 
es can be 
| the pro- 
each sub 
Z values 
finds the 
e lines, a 
where the 
  
maximum correlation is assumed as the approximate coordinate 
of the ALSM. 
We use, a computer with 1 GH cpu and 128 MB RAM. With 
this computer matching of one pixel only on two images takes 
about 3 seconds. For precise surface reconstruction, we want to 
match all pixels in the interest region. And their number is at 
least 20 000. With this equipment matching these points only on 
two images requires about 27 hours. So we couldn't still obtain 
the external face. But in future works we plan to buy a parallel 
processor. Then we think we can overcome the speed problem. 
We don't have pattern projector either. So during the matching 
process, on the image areas where the textures are less, false 
matchings occur. After we have overcome these problems, we 
hope to have good results. 
5. APPLICATIONS ON FINAL MODEL 
In this section, we will explain our software with sample images 
produced by the software. In medical applications, by using 3D 
models and other processing tools, it is desired that the users 
can visualize the models from every location with desired col- 
ours. And both for diagnosis and treatment, various imaging 
opportunities should alse be supplied. When necessary, users 
must clip, cut, resample etc. the models and must see the parts 
of the interest body that they want. In this section we'll give the 
very useful examples produced by our software. In spite of 
these examples, there are currently many other tools, which can 
be used for diagnosis too. 
In Figure 1, segmented brain image is shown. This segmenta- 
tion has performed by the methods explained in section 3. 
d 
  
  
  
  
  
Figure 1. Segmentation of brain image 
Figure 2 shows, segmented image is superimposed to original 
image. 
  
  
  
  
  
  
  
Figure 2. Segmented image superimposed to original. 
During segmentation, user can control the segmentation by 
superimposing it to original. By superimposing, when desired, 
opacity and colour classification can also be performed for 
volume rendering. 
In Figure 3, 3D surface model of the brain which produced with 
the segmented brain images is shown. In the figure, vertex 
coordinates of the surface are listed in a tubular format. 
  
Figure 3. Surface model of the brain 
In Figure 4, brain tumour surface model is shown. In the figure, 
left image is tumour surface without smoothing and on the 
right, smoothed tumour surface. Smoothing errors can be seen 
in Figure 5 both graphically and with the error values on each 
vertex in tubular format. On the black background window, 
some of the segmented tumour image slices is seen. Tumour 
surface model has obtained with this segmented tumour slices. 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Figure 5. Smoothing errors of tumour surface model 
Figure 6 shows the brain with the tumour. Tumour is on the 
lower right corner of the brain with red colour. Tumour surface 
-253- 
  
PEUT NE ass