What kind of features can be extracted from knee MRI?

Hi all, Do you know what features can be extracted from knee MRI for cartilage detection? I want to extract some features to classify voxels. If there is some papers for feature extraction, I will be thankful if you share.

Answer 1

[Star Strider]

The literature on that is not extensive, but available if your library has access to the journals. See my PubMed search for one article, and see the ‘Similar Articles’ section as well.

[Sara Salimi]

Dear Star Strider many thanks for your answer, and for the paper. I am beginner in Image Processing and I need to extract the following features. All features except the position are calculated at three different scales (0.65, 1.1, and 2.5):

• local 3D geometric features (3-jet)
• intensities from convolution with Gaussian derivatives with 3 scales (0.65, 1.1, and 2.5)
• pixels intensity
• pixel positions in the image
• eigenvalues and eigenvectors of the Hessian
• structure tensor eigenvalues and eigenvectors combining three different scales. All derivatives are achieved by convolution with Gaussian derivatives, defined as where where G is a Gaussian, D the differentialoperator and σ is the scale.

Please help me how to extract these features? or please introduce me some resources to learn about these features of images. Is there any books for beginners?

Thanks in advance

[Image Analyst]

See Steve Eddins book on Image Processing with Matlab, available from online bookstores such as Amazon.com.

http://matlab.wikia.com/wiki/FAQ#Can_you_program_up_the_algorithm_in_this_article_for_me_and_explain_it_to_me.3F

Answer 2

[Walter Roberson]

The number of different features that can be computed is immense. Even at the very simplest, you might choose to extract k pixels out of n (there are (n choose k) ways of doing that), but you might also choose to order those in any of the k! different ways. The total number of possibilities gets big fast -- like 3554627472076 possibilities for only 16 total pixels. See http://oeis.org/A000522

But as well as just selecting some subset of pixels, you might choose to do some calculation on the subset. For example, you might take sin() of them. Or dct() . Or blockwise local maximum. Or matrix multiplication that projects from dimension (p x q) to dimension (r x s).

You need to understand what, mathematically a "feature" is. Mathematically, a "feature" is any function that takes your array of pixels and produces a vector output, typically of a smaller size, and possibly not of the same data type. This involves the mathematical meaning of "function" as opposed to "formula": it does not have to be "computable", it just has to be something that could be looked up in a huge table. Playing Fizzbin with your matrix of pixels is entirely acceptable for the purpose of defining what a "feature" is.

What features can be extracted? Just about anything you can think of that can be computed in a finite time.

The question you need to ask is what features are useful to compute. And that is still very much under study. See Star Strider's response for some relevant articles.

[Sara Salimi]

Thanks for your answer