Two thoughts (based on my 2 minutes of understanding of this problem from
reading your email):
1. It seems that the 0 values are important and indicate soft tissue. I am not
sure why you couldn't use these discriminators. If they are 0 in all samples
then they won't impact the classifier, if 0 in some then maybe you can classify
by these values.
2. It sounds like the data are images and I wonder if work on fnctional data
analysis would be useful. Search the internet on RAMSAY STATISTICS MONTREAL
and you will get to his web page and a link to the book on functional data
analysis.
DO YOU WANT TO COME TO ST LOUIS NEXT JULY 5-7 FOR THE 2008 ANNUAL MEETING OF
THE CLASSIFICATION SOCIETY AND TALK ABOUT YOUR WORK????
Bill
"Audette, Michel" <[EMAIL PROTECTED]> wrote: v\:*
{behavior:url(#default#VML);} o\:* {behavior:url(#default#VML);} w\:*
{behavior:url(#default#VML);} .shape {behavior:url(#default#VML);}
st1\:*{behavior:url(#default#ieooui) } Hi Bill,
thanks for your kind reply. One feature is CT data of the head, while the
second is a sheetness measure, based on a paper by Descôteaux
Bone enhancement filtering: application to sinus bone segmentation and
simulation of pituitary surgery
www-sop.inria.fr/odyssee/team/Maxime.Descoteaux/docs/descoteaux_miccai05.pdf
It detects locally curviplanar tissues at various scales, and in fact the
scale coinciding with the maximum sheetness value is also stored.
This sheetness data is a volume image, and because of the way it is output,
using a minc volume where reals are stored as shorts with a scaling factor, any
value less than 0.001 is simply output as null in these volumes. So much of
soft tissue in the head, except skin which is in fact somewhat sheetlike, as
well as background, has a null value, where cranial bones have a high value.
If I use a minimally supervised classification algorithm, I have to do it in
a manner where one of the features has zero values for some tissues, and
non-zero values for others, unless I replace those zero values by a synthetic
value for mean and variance, in keeping with that .001 threshold.
What do you think?
Cheers,
Michel
Michel Audette, Ph.D.
Innovation Center Computer Assisted Surgery (ICCAS)
Philipp-Rosenthal-Strasse 55
Leipzig, Germany
Phone: ++49 (0) 341 / 97 - 1 20 13
Fax: ++49 (0) 341 / 97 - 1 20 09
---------------------------------
From: Classification, clustering, and phylogeny estimation [mailto:[EMAIL
PROTECTED] On Behalf Of William Shannon
Sent: July 25, 2007 6:59 PM
To: [email protected]
Subject: Re: using features which may be null for some classes
To you explain the data a little more?
Bill Shannon
"Audette, Michel" <[EMAIL PROTECTED]> wrote:
Dear all,
I am interested in implemented a semi-supervised clustering method, i.e.:
making use of a small set of training points, to classify tissues of the head
visible in CT data. I would like not only to use CT intensity as a feature, but
a measure of sheet-like structure that correlates with thin bone, and may
assist the detection of thin bone structures that are otherwise undiscernible
from soft tissue, due to partial volume effects that blurr intensities
together. However, this latter feature, sheetness, produces a null value for
most tissue classes.
Can anyone suggest a means of integrating two features together, CT and
sheetness, in a clustering algorithm, given that one of them appears null for
several classes?
Best regards,
Michel Audette, Ph.D.
Innovation Center Computer Assisted Surgery (ICCAS)
Philipp-Rosenthal-Strasse 55
Leipzig, Germany
Phone: ++49 (0) 341 / 97 - 1 20 13
Fax: ++49 (0) 341 / 97 - 1 20 09
----------------------------------------------
CLASS-L list.
Instructions: http://www.classification-society.org/csna/lists.html#class-l
---------------------------------------------- CLASS-L list. Instructions:
http://www.classification-society.org/csna/lists.html#class-l
---------------------------------------------- CLASS-L list. Instructions:
http://www.classification-society.org/csna/lists.html#class-l
----------------------------------------------
CLASS-L list.
Instructions: http://www.classification-society.org/csna/lists.html#class-l
