[Insight-users] ITK ROAD MAP 2005-2006 : Call for feedback

Tue May 31 12:11:16 EDT 2005

Miller, James V (Research) wrote:

>Zachary, 
>
>Is the current itk::MRFImageFilter not sufficient?
>
I think that implementation only uses ICM.

>  Does it need to be
>refactored to accomadate different MRF algorithms?
>  
>
It would be valuable to have some other options e.g. mean field, 
simulated annealing.

>I only used the MRFImageFilter once. It seemed to perform as I would have
>expected (from reading some of the literature).
>
>One thing I think ITK probably needs are more techniques for learning the 
>pdf's for each class of material.
>
>Jim
>
>
>
>
>-----Original Message-----
>From: insight-users-bounces+millerjv=crd.ge.com at itk.org
>[mailto:insight-users-bounces+millerjv=crd.ge.com at itk.org]On Behalf Of
>Zachary Pincus
>Sent: Saturday, May 28, 2005 7:37 PM
>To: ITK mailing
>Subject: Re: [Insight-users] ITK ROAD MAP 2005-2006 : Call for feedback
>
>
>Hi all,
>
>After looking over the 2005-2006 ITK roadmap, I've also got a couple of 
>questions/comments on the machine learning aspects.
>
>Specifically, to what ends are classification algorithms (e.g. gaussian 
>mixture models, k-nearest neighbors, putative neural networks or SVMs) 
>present in ITK? It strikes me that one major use of such algorithms in 
>medical imaging is for classification of image pixels into various 
>tissue types, e.g. grey matter vs. white matter.
>
>If this is the case, I would think that adding Markov Random Field 
>capabilities to ITK would be a big win. Basically, MRFs allow users to 
>add priors about the *spatial* distribution of various pixel types into 
>the classification process. For example, a single isolated pixel 
>initially labeled as "grey matter" in a blob of white matter might 
>(depending on the priors) be considered an unlikely configuration and 
>thus be re-labeled in the final MRF configuration. Such spatial 
>considerations are ignored by traditional classifiers.
>
>Because spatial information is so important, and MRFs are a relatively 
>easy way to add simple spatial priors, they have become quite popular 
>in the image processing literature. I think filters to estimate the MAP 
>MRF given an input "label images" (e.g the results of pixel-wise 
>classification) would be a very valuable addition, especially if the 
>stable of pixel classification methods in ITK is to expand.
>
>Now, I haven't described in too much detail how MRF models actually 
>work. A Google Scholar search for "Markov random field image" will show 
>the breadth of utilization of MRFs in the imaging literature. Here is a 
>good introduction to MRF segmentation, with specific reference to MRI 
>images:
>Segmentation of brain MR images through a hidden Markov random field 
>model and the expectation-maximization algorithm.
>Y Zhang, M Brady, S Smith - IEEE Trans Med Imaging, 2001
>http://www.cvmt.dk/~hja/teaching/cv/HMRF_EM_BRAIN.pdf
>
>I would be happy to discuss at (much) more length how a MRF 
>"classification cleanup" filters could be implemented in ITK, if there 
>is any interest in these methods.
>
>Zach
>
>
>
>On May 27, 2005, at 2:46 PM, Lino Ramirez wrote:
>
>  
>
>>Hi Luis and ITK Users/Developers,
>>
>>I had a brief look at the ITK roadmap 2005-2006. It looks quite
>>impressive. I cannot wait until having available all these tools in one
>>single package ;-)
>>
>>I have some small comments/questions about functionalities I would 
>>like to
>>see in the toolkit.
>>
>>I noticed that Neural Networks will be added to the toolkit. Are there 
>>any
>>plans for adding a Support Vector Machines (SVM) [1] implementation? 
>>SVM
>>have been used successfully in a variety of applications that could be 
>>of
>>interest to the ITK community (see [2] for some sample applications).
>>Moreover, it is always good to have a machine learning approach that is
>>similar to the neural networks in architecture but that uses a 
>>different
>>learning strategy. In this way, one could try the two of them and
>>determine which one is more appropriate for a particular dataset.
>>Sometimes, in datasets in which the neural networks fail the SVM 
>>succeed
>>and vice versa.
>>
>>Are there any plans (even in the very long term) to add support for 
>>Fuzzy
>>Sets [3], Fuzzy Geometry [4], and Fuzzy Spatial Relations [5] between
>>objects in an image. I think these concepts would be invaluable in the
>>future of medical image analysis. For example, when we want to measure
>>geometric properties in objects in an image, we find that generally the
>>objects are not crisply defined (due to errors during the segmentation,
>>errors in the acquisition of the image, or errors in the definition of 
>>the
>>object –where do the ribs start and the vertebrae end in a spine 
>>X-ray).
>>In this case, fuzzy geometry could be used to compute the object
>>properties. Another example would be in the identification of objects 
>>in
>>the images. For instance, in the internal brain structures, the right
>>caudate nucleus should be closer to the right lateral ventricle than to
>>the left lateral ventricle. Fuzzy spatial relations with the help of 
>>fuzzy
>>logic [6] could be used to develop a system that makes use of that 
>>piece
>>of information to identify right lateral ventricle.
>>
>>Well, those are my two picks ;-)
>>
>>I am looking forward to any comment
>>
>>Take care
>>
>>Lino
>>
>>[1] C. Cortes and V. Vapnik, "Support-Vector Networks," Machine 
>>Learning,
>>vol. 20, pp. 273-297, 1995
>>[2] http://www.clopinet.com/isabelle/Projects/SVM/applist.html
>>[3] L.A. Zadeh, "Fuzzy sets," Information and Control, vol. 8, pp. 
>>38-352,
>>1965
>>[4] A. Rosenfeld, "Fuzzy geometry: An updated overview," Information
>>Sciences, vol. 110, pp. 127-133, 1998
>>[5] I. Bloch, "Fuzzy spatial relationships for image processing and
>>interpretation: a review," Image and Vision Computing, vol. 23, pp.
>>89-110, 2005
>>[6] L.A. Zadeh, "Outline of a new approach to the analysis of complex
>>systems and decision processes," IEEE Transactions on Systems, Man, and
>>Cybernetics, vol. SMC-3, no. 1, pp. 28-44, 1973
>>
>>    
>>
>>>A first draft of the road map for ITK development/maintenance has
>>>been crafted for the period of September 2005 - September 2006.
>>>
>>>
>>>You will find this draft as a link to the Oversight Committee page
>>>
>>>http://www.itk.org/Wiki/ITK_Oversight_Committee
>>>
>>>
>>>More specifically at
>>>
>>>
>>>http://www.itk.org/Wiki/ITK_Roadmap_2005_2006
>>>
>>>
>>>The purpose of this road map is to plan for features and 
>>>functionalities
>>>to be included in ITK in the near/medium term (1 to 2 years).
>>>
>>>The addition of these features should make of ITK a better tool for
>>>supporting your efforts in medical research, and development of 
>>>medical
>>>applications.
>>>
>>>The road map also includes the maintenance tasks to be undertaken in
>>>ITK. This may involve refactoring of classes, deprecation of classes,
>>>additional testing, additional coverage, improvements on tutorials and
>>>so on.
>>>
>>>
>>>Please let us know of the features that you would like to see in ITK
>>>in the upcoming future, and what points of the toolkit you consider
>>>that can be improved in order to better server the community.
>>>
>>>
>>>
>>>Thanks
>>>
>>>
>>>
>>>Luis
>>>
>>>      
>>>
>>_______________________________________________
>>Insight-users mailing list
>>Insight-users at itk.org
>>http://www.itk.org/mailman/listinfo/insight-users
>>
>>    
>>
>
>_______________________________________________
>Insight-users mailing list
>Insight-users at itk.org
>http://www.itk.org/mailman/listinfo/insight-users
>_______________________________________________
>Insight-users mailing list
>Insight-users at itk.org
>http://www.itk.org/mailman/listinfo/insight-users
>  
>

-- 
Simon K. Warfield, Ph.D. warfield at bwh.harvard.edu Phone:617-732-7090
http://www.spl.harvard.edu/~warfield           FAX:  617-582-6033
Associate Professor of Radiology,          Harvard Medical School
Director, Computational Radiology Laboratory
Thorn 329, Dept Radiology,  Brigham and Women's Hospital 
75 Francis St, Boston, MA, 02115