[Insight-developers] CannySegmentationLevelSetImageFilter

Miller, James V (Research) millerjv at crd . ge . com
Fri, 21 Nov 2003 14:31:11 -0500


This message is in MIME format. Since your mail reader does not understand
this format, some or all of this message may not be legible.

------_=_NextPart_001_01C3B066.04C548E2
Content-Type: text/plain;
	charset="iso-8859-1"

I mentioned on the tcon this afternoon that we were having some problems
with the CannySegmentationLevelSetImageFilter.  If we used a positive weight
for the advection weight, our segmentations would contract whereas if we
used a negative weight for the advection weight our segmentations would
expand (like we wanted).  
 
Lydia mentioned that it may be due to how the seed image is being used.
Once again, Lydia is right. We were passing in a binary image (the output of
another segmentation algorithm) as the seed image.  The seed image had 255
for the object and 0 for the background.  The SparseFieldLevelSetImageFilter
takes this seed image and generates an initial level set by assigning pixels
above a specified IsoValue as being outside the object and pixels below a
specified IsoValue as being inside the object.  So while I thought my
segmentation was shrinking, it was actually growing.  It was just that my
"object" was actually my background.
 
So.  Here are my suggestions.  One, we should be very clear that the inside
of objects have negative level set values and the outside of objects have
positive level set values.  Two, we should change the documentation for
*SegmentationLevelSetFilters so that it is clear that the seed image follows
the same conventions for objectness as the levelsets, i.e. values less than
the specified IsoValue are interior to the object, values greater than the
specified IsoValue are exterior to the object.  I am not sure whether we
want to go so far as to say the seed image is an initial level set (which
does not really have to be) or whether we can just say that it follows the
same conventions as level sets.
 
Finally, I am not sure when we changed the levelsets to be consistent with
negative interiors and positive exteriors but I regardless of when it was
done, it looks like the User Guide is out of date.  We'll need to propagate
this into the User Guide and perhaps put out a list of "errata" on the web
site.
 
Any comments?
 
 

Jim Miller 
_____________________________________
Visualization & Computer Vision
GE Research
Bldg. KW, Room C218B
P.O. Box 8, Schenectady NY 12301

millerjv at research . ge . com <mailto:millerjv at research . ge . com> 

james . miller at research . ge . com
(518) 387-4005, Dial Comm: 8*833-4005, 
Cell: (518) 505-7065, Fax: (518) 387-6981 

 

------_=_NextPart_001_01C3B066.04C548E2
Content-Type: text/html;
	charset="iso-8859-1"

<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.0 Transitional//EN">
<HTML><HEAD>
<META HTTP-EQUIV="Content-Type" CONTENT="text/html; charset=iso-8859-1">


<META content="MSHTML 6.00.2800.1264" name=GENERATOR></HEAD>
<BODY>
<DIV><SPAN class=013561719-21112003><FONT size=2>I mentioned on the tcon this 
afternoon that we were having some problems with the 
CannySegmentationLevelSetImageFilter.&nbsp; If we used a positive weight for the 
advection weight, our segmentations would contract whereas if we used a negative 
weight for the advection weight our segmentations would expand (like we 
wanted).&nbsp; </FONT></SPAN></DIV>
<DIV><SPAN class=013561719-21112003><FONT size=2></FONT></SPAN>&nbsp;</DIV>
<DIV><SPAN class=013561719-21112003><FONT size=2>Lydia mentioned that it may be 
due to how the seed image is being used.&nbsp; Once again, Lydia is right. We 
were passing in a binary&nbsp;image (the output of another segmentation 
algorithm) as the seed image.&nbsp; The seed image had 255 for the object and 0 
for the background.&nbsp; The SparseFieldLevelSetImageFilter takes this seed 
image and generates an initial level set by assigning pixels above a specified 
IsoValue as being outside the object and pixels below a specified IsoValue as 
being inside the object.&nbsp; So while I thought my segmentation was shrinking, 
it was actually growing.&nbsp; It was just that my "object" was actually my 
background.</FONT></SPAN></DIV>
<DIV><SPAN class=013561719-21112003><FONT size=2></FONT></SPAN>&nbsp;</DIV>
<DIV><SPAN class=013561719-21112003><FONT size=2>So.&nbsp; Here are my 
suggestions.&nbsp; One, we should be very clear that the inside of objects have 
negative level set values and the outside of objects have positive level set 
values.&nbsp; Two, we should change the documentation for 
*SegmentationLevelSetFilters so that it is clear that the seed image follows the 
same conventions for objectness as the levelsets, i.e. values less than the 
specified IsoValue are interior to the object, values greater than the specified 
IsoValue are exterior to the object.&nbsp; I am not sure whether we want to go 
so far as to say the seed image is an initial level set (which does not really 
have to be) or whether we can just say that it follows the same conventions as 
level sets.</FONT></SPAN></DIV>
<DIV><SPAN class=013561719-21112003><FONT size=2></FONT></SPAN>&nbsp;</DIV>
<DIV><SPAN class=013561719-21112003><FONT size=2>Finally, I am not sure when we 
changed the levelsets to be consistent with negative interiors and positive 
exteriors but I regardless of when it was done, it looks like the User Guide is 
out of date.&nbsp; We'll need to propagate this into the User Guide and perhaps 
put out a list of "errata" on the web site.</FONT></SPAN></DIV>
<DIV><SPAN class=013561719-21112003><FONT size=2></FONT></SPAN>&nbsp;</DIV>
<DIV><SPAN class=013561719-21112003><FONT size=2>Any 
comments?</FONT></SPAN></DIV>
<DIV><SPAN class=013561719-21112003><FONT size=2></FONT></SPAN>&nbsp;</DIV>
<DIV><SPAN class=013561719-21112003><FONT size=2></FONT></SPAN>&nbsp;</DIV>
<DIV>
<P style="MARGIN: 0in 0in 0pt"><B><SPAN 
style="COLOR: navy; FONT-FAMILY: 'Comic Sans MS'">Jim Miller</SPAN></B> 
<BR><B><I><SPAN 
style="FONT-SIZE: 10pt; COLOR: red; FONT-FAMILY: Arial">_____________________________________</SPAN></I></B><BR><EM><SPAN 
style="FONT-SIZE: 7.5pt; COLOR: black; FONT-FAMILY: Arial">Visualization &amp; 
Computer Vision</SPAN></EM><I><SPAN 
style="FONT-SIZE: 7.5pt; COLOR: black; FONT-FAMILY: Arial"><BR><EM>GE 
Research</EM><BR><EM>Bldg. KW, Room C218B</EM><BR><EM>P.O. Box 8, Schenectady NY 
12301</EM><BR><BR></SPAN></I><EM><U><SPAN 
style="FONT-SIZE: 7.5pt; COLOR: blue"><A 
href="mailto:millerjv at research . ge . com">millerjv at research . ge . com</A></SPAN></U></EM></P>
<P style="MARGIN: 0in 0in 0pt"><EM><U><SPAN 
style="FONT-SIZE: 7.5pt; COLOR: blue">james . miller at research . ge . com</SPAN></U></EM><BR><I><SPAN 
style="FONT-SIZE: 7.5pt; COLOR: black; FONT-FAMILY: Arial">(518) 387-4005, Dial 
Comm: 8*833-4005, </SPAN></I><BR><I><SPAN 
style="FONT-SIZE: 7.5pt; COLOR: black; FONT-FAMILY: Arial">Cell: (518) 505-7065, 
Fax: (518) 387-6981</SPAN></I> </P></DIV>
<DIV>&nbsp;</DIV></BODY></HTML>

------_=_NextPart_001_01C3B066.04C548E2--