ImageJ PCA
Locked
 
|
Dear all,
I have a question in regards to the ImageJ plugins for PCA. I am trying to extract the principal components of a feature set that comprises of 56 features. I used the ImageJ plugin for calculating the PCA for a stack of images and in the end i get a graph with the eigen values for the respective images. Initially i used 14 stacks to calculate the PCA, of them i got the eigen values maximum for the images i loaded first in the stack. The images that were loaded in the end returned 0 or minimal eigen values. I tried to reverse the process to validate if the PCA calculated is correct by loading the same images which returned 0 eigen values in the previous exercise first, and loaded those images that returned the maximum eigen values in the end. This time i got the opposite result, the images loaded first had the maximum eigen values and the images loaded in the end returned 0 eigen values. I assume the images that are loaded first, their eigen values somehow always return the maximum value. Also there is no projection graph of the PCA. Is the application of PCA in the ImageJ plugins useful to my case in which i want to reduce the feature set? Thanks in advance. Regards, Ali. |
|
|
On Wednesday 11 April 2007 18:40:30 aszia787 wrote:
> Dear all, > > I have a question in regards to the ImageJ plugins for PCA. I am trying to > extract the principal components of a feature set that comprises of 56 > features. I used the ImageJ plugin for calculating the PCA for a stack of > images and in the end i get a graph with the eigen values for the > respective images. Initially i used 14 stacks to calculate the PCA, Maybe I do not understand. If you have 56 variables, where does the 14 (image stack?) come from? I would have expected that each slice represents a different variable when one submits the stack to the PCA (here I am talking of the BIJ plugin called PCA). Cheers, G. |
|
|
Hi Gabriel,
Thanks for your reply. By 14 i mean that instead of submitting the entire 56 images, i assumed that i had a feature set of 14 images(not giving a stack of 56 images as its too tedious a task :)). It returned me a stack of 14 eigenimages after the PCA. I just want to extract the principal components. Example, if 14 images were submitted, i wanted to know the eigen values of each of the 14 images, and would have selected the first 3 highest eigen values of the 14 images to have the pricipal component and ignore the rest of the features. Is it possible with the BIJ plugin. Thanks alot. Regards, Ali
|
|
|
On Friday 13 April 2007 13:01:07 aszia787 wrote:
> I just want to extract the principal > components. Example, if 14 images were submitted, i wanted to know the > eigen values of each of the 14 images, and would have selected the first 3 > highest eigen values of the 14 images to have the pricipal component and > ignore the rest of the features. Is it possible with the BIJ plugin. Aren't these the greyscale values of the new stack (eigenimages as reported to the console)? These are returned as 32 bit. First image is the first component and so on? As you go along the new stack there should be less explanation the variance (please correct me if I got this wrong). Regards, G. |
|
|
Gabriel,
First of all i'm so thankful to you for your patience in answering my questions, i highly appreciate it. Yes, that's true, i get 32 bit images in the stack and i assumed that the first image i get has the maximum eigenvalue. I'll tell you what i am trying to achieve from the PCA. As mentioned earlier, i have a set 56 features/images and i want to extract say the highest 4 eigenvalues. After getting the 4 eigenvalues i will add only the pricipal component planes(4 images) to get the final feature image. I will try and give you an example. Assume we have 5 images A,B,C,D and E and we pass these images in a stack and extract the PCA. The PCA that we get is in the form of 32 bit images with the highest variance being the first component and so on. My question is, is the first eigenimage that we get, is that the eigneimage of image A or image B? i guess it doesn't tell me that the first image returned is the eigen image of which original feature/image as i have to add the original features with the maximum eigenvalue. I hope i was able to clear the question. If you have any questions, please feel free to ask, i will try to clarify again if needed. Once again thank you so much for your interest. Regards, Ali.
|
|
|
On Friday 13 April 2007 14:38:40 aszia787 wrote:
> My question is, is the first eigenimage that we > get, is that the eigneimage of image A or image B? i guess it doesn't tell > me that the first image returned is the eigen image of which original > feature/image as i have to add the original features with the maximum > eigenvalue. I am no expert in PCA , but as I understand it, the plugin calculates new combinations of the original images that explain the variance of the data and puts them in decreasing order of "variance explanation". That is returned, I believe, in the eigenvalues spectrum graph (but please correct me if I am in the wrong track). I do not think that the plugin gives you information about "what component corresponds to mixtures of which images" like is reported in some stats programs like SPSS or Minitab. There is some information with references here (but perhaps you already know this): http://bij.isi.uu.nl/pca.htm Have you consulted Michael Abràmoff (the author of the plugin)? Cheers, G. |
|
|
Thanks for your reply. I kindaa agree what you have analyzed. It doesn't tell the eigen value of the original image. I did try to contact Michael Abràmoff, he has not responded me yet. Michael if you are reading this, please give your feeback. Is there any other free software that you think would solve my problem. SPSS is not a free software i belive, dunno much about Mintlab. Any other sugesstions you feel could be a replacement for PCA to remove the redundant features and extract the uncorrelated features?
Regards, Ali.
|
|
|
Jim Burger presented an ImageJ-based program for PCA in the ImageJ
Conference 2006, Luxemburg. It's worth to check out. Albert > problem. SPSS is not a free software i belive, dunno much > about Mintlab. Any other sugesstions you feel could be a replacement for PCA > to remove the redundant features and extract the uncorrelated features? > > Regards, > > Ali. |
|
|
Hi Albert,
Thanks for your feedback. I have been trying to locate the ImageJ plugin of Jim Burger but i guess the website of the plugins of the conference is not working somehow. Is it possible for you to send the plugin or can you refer me to a specific site from where i could download the plugin. Thanks in advance. Regards, Ali.
|
|
|
I am not sure, but I remember that the program of Jim was not an IJ
plugin. Try to contact JIm at : http://www.burgermetrics.com/ Gary. www.gcsca.net On Apr 14, 2007, at 6:04 PM, aszia787 wrote: > Hi Albert, > > Thanks for your feedback. I have been trying to locate the ImageJ > plugin of > Jim Burger but i guess the website of the plugins of the conference > is not > working somehow. Is it possible for you to send the plugin or can > you refer > me to a specific site from where i could download the plugin. > Thanks in > advance. > > Regards, > > Ali. > > Albert Cardona wrote: >> >> Jim Burger presented an ImageJ-based program for PCA in the ImageJ >> Conference 2006, Luxemburg. It's worth to check out. >> >> Albert >> >>> problem. SPSS is not a free software i belive, dunno much >>> about Mintlab. Any other sugesstions you feel could be a >>> replacement for >>> PCA >>> to remove the redundant features and extract the uncorrelated >>> features? >>> >>> Regards, >>> >>> Ali. >> >> > > -- > View this message in context: http://www.nabble.com/ImageJ-PCA- > tf3560478.html#a9994452 > Sent from the ImageJ mailing list archive at Nabble.com. > > |
|
|
In reply to this post by aszia787
Hi,
I did not understand completely the nature of the problem raised. But, if this can be useful, can I suggest you to try "my" PCA (and CA: Correspondence Analysis, another variant to Multivariate Statistical Analysis) version, available at: http://www.univ-reims.fr/INSERM514/ImageJ I do not claim that "my" version is superior to the other ones. But, may be, a different presentation of things can help to solve or understand the problem. Hope it helps. Noel Bonnet |
|
|
Hi Noel,
Thanks for your feedback. I think i have used your plugin, just wanted to ask you if i submit 5 images, say A, B, C, D and E, would your program tell me the eigen values returned of each original submitted image. I need to extract the principal component plane of each submitted image and keep the maximum eigen values of the submitted images. Thanks for your feedback. Regards, Ali.
|
«
Return to ImageJ
|
1 view|%1 views
| Free forum by Nabble | Edit this page |