I like the MRF for its ability to apply spatial coherence to improve the segmentation, but I'm not entirely sure if it'll work with my specific problem (or at least the way I have it set up). My problem is that 3 of my classes are boundary classes (i..e stool-tissue vs. just stool). <div>
<br></div><div>As air (-1000 HU) transitions to stool (1000 HU) it crosses the range for tissue (0 HU). Because of this, the intensity of stool-air voxels closely mimics tissue and confuses the intensity average used in the distance metric of the MRF.</div>
<div><br></div><div>Can the MRF deal with "unclassified" voxels? Or does it only improve an existing, complete segmentation?</div><div><br></div><div>How can I avoid treating the boundaries as distinct classes? Can I use a mask in some way? <br>
<br><div class="gmail_quote">On Sat, May 21, 2011 at 2:47 PM, Luis Ibanez <span dir="ltr"><<a href="mailto:luis.ibanez@kitware.com">luis.ibanez@kitware.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex;">
Hi Neil,<br>
<br>
This refinement of the classification looks like a natural task for<br>
Markov Random Fields, where you can express the expectation<br>
of neighborhoods between the different classes in the image.<br>
<br>
You may want to look at:<br>
<a href="http://www.itk.org/Doxygen/html/classitk_1_1MRFImageFilter.html" target="_blank">http://www.itk.org/Doxygen/html/classitk_1_1MRFImageFilter.html</a><br>
<br>
<br>
Luis<br>
<br>
<br>
---------------------------------------------------------------------------<br>
<div><div></div><div class="h5">On Fri, May 20, 2011 at 4:47 AM, Neil Panjwani <<a href="mailto:paniwani@gmail.com">paniwani@gmail.com</a>> wrote:<br>
> Hi,<br>
> I need some help with a CT imaging artifact. I'm working on a stool<br>
> subtraction project of the colon in which the stool is tagged near 700 HU<br>
> (bright), air is -1025 HU (dark) and tissue is grey in the range of -300-400<br>
> HU (depending on the proximity to stool).<br>
> I'm trying to separate stool-air boundaries from stool-tissue-air<br>
> boundaries. The problem arises because of the partial volume effect of CT<br>
> imaging. At the stool-air boundary, the intensity is blurred and mimics<br>
> tissue as the intensity goes from very negative to very positive. The same<br>
> occurs during a stool-tissue-air boundary, however the change is less abrupt<br>
> due to the inner tissue layer.<br>
> I've tried using a connected component from known tissue but I erroneously<br>
> connect stool-air as tissue in the process. I've tried intensity and<br>
> gradient magnitude thresholds but the intensity-gradient relationship for<br>
> the two cases are nearly identical. The gradient is very sensitive to my<br>
> choice of sigma; values too large lose the tissue altogether and values too<br>
> low keep too much noise. I've also tried directional gradients, moving<br>
> outward from air (dark), but again the choice of sigma makes it extremely<br>
> sensitive.<br>
> I have also tried using local Haralick texture features (in the OTB<br>
> toolbox), but haven't seen good results. I've only tried a few offset pairs.<br>
> I try to avoid computing all offsets and averaging because it is incredibly<br>
> slow to generate all the GLCM histograms.<br>
> I've attached the original input and my voxel map, based on simple intensity<br>
> and gradient thresholds, as well as zoomed in versions illustrating the<br>
> problem. Here, orange represents tissue, black is stool, blue is air, and<br>
> white is unclassified. As can be seen in map_zoom.png, tissue which is<br>
> bordering both stool and air in the center of the image is left<br>
> unclassified.<br>
> How can I distinguish these two boundaries in order to recover the<br>
> unclassified tissue?<br>
</div></div>> _____________________________________<br>
> Powered by <a href="http://www.kitware.com" target="_blank">www.kitware.com</a><br>
><br>
> Visit other Kitware open-source projects at<br>
> <a href="http://www.kitware.com/opensource/opensource.html" target="_blank">http://www.kitware.com/opensource/opensource.html</a><br>
><br>
> Kitware offers ITK Training Courses, for more information visit:<br>
> <a href="http://www.kitware.com/products/protraining.html" target="_blank">http://www.kitware.com/products/protraining.html</a><br>
><br>
> Please keep messages on-topic and check the ITK FAQ at:<br>
> <a href="http://www.itk.org/Wiki/ITK_FAQ" target="_blank">http://www.itk.org/Wiki/ITK_FAQ</a><br>
><br>
> Follow this link to subscribe/unsubscribe:<br>
> <a href="http://www.itk.org/mailman/listinfo/insight-users" target="_blank">http://www.itk.org/mailman/listinfo/insight-users</a><br>
><br>
><br>
</blockquote></div><br></div>