[Insight-users] Approximate segmentation of very high intensity region.

[Xi Liang]

Hi all,

I have this problem that to approximated segment the enhanced tumor region
in subtraction images from post- to pre-contrast images of breast DCEMRI
images without any seed or threshold as input. I attach the central slice
with this email.

In the subtraction image, the intensity range of enhanced tumor (region to
be segmented) is 400-600 and the other tissue/background is 0-200.

I am not quite familiar with segmentation/classification methods and just
doing some trivial analysis on what my problem is. Could someone comment on
the analysis and please give some suggestions on which
segmentation/classification method would be better to be used in this
problem.

I can think two ways to do this one is thresholding-based method and the
other is region grow.

1. I will need to automatically find a (higher) threshold hoping that the
mask of tumor region will not include too much background pixels. I tried
Ostu method in ITK, however the result mask include too much background
pixels. The image Ostu.png is attached. I think a better method should be an
algorithm that assign pixels/intensity into two clusters where the
difference of means of two clusters is maximum. If this is right, I think
KMeans method could be a better choice which I have not tried yet.

2. If using region grow method, I will need to find the seed point and then
apply it on ConnectThresholding filter. However the problem is though
enhanced tumor region usually have a higher intensity, but not always has
the maximum intensity in the subtraction image. (Some artifact might also
highly enhanced).

3. I also think of other segmentation method like waterheld or level-set
methods. They might not suitable for this case. My understanding of these
two method through the SegmentationOverview.pdf is that they are more useful
on edge detection or local segmentation.

My feeling is to find a threshold for the image to do segmenation, can
someone suggest a suitable classification/segmentation method? I am
currently trying KMeans.

Kind regards,

Xi http://old.nabble.com/file/p30052326/subtraction.png 
http://old.nabble.com/file/p30052326/Ostu.png 

-- 
View this message in context: http://old.nabble.com/Approximate-segmentation-of-very-high-intensity-region.-tp30052326p30052326.html
Sent from the ITK - Users mailing list archive at Nabble.com.