Hello Shaima, which kind of image processing do you want?

Dear Jose,

Thank  you  for  replying. 

Image enhancement  and segmentation. 

Ok, so let'see.

Currently I am working with machine learning and deep learning techniques to segment medical images( subcortical brain structures on MRI images).

Let's say that in a coarse view, these techniques need an input vector of samples that are composed by (ideally) discriminative features. This vector is used in a first stage to train the classifier. During the classification stage, the samples to be classified are composed by the same features, thus the classifier can take the proper decision. This is of course a very very weak explanation, but I think this can help to have a first contact with these techniques.

Regarding the features, a huge range of options can be used. At least for segmentation, from basic texture values, like the value of the pixel under examination and/or some relations with its neighborhood, to some more complex features, such as Gabor filter values, Gradient or wavelet decomposition values, for example. Additionally, some spatial information (such as pixel location) or a priori information (like probablity maps) can be used.

As initial example, imagine that you have a white and black image that you want to segment, only based on gray values and into two classes (A and B).

Your input vector will look like:

          sample 1 [ class A grayValue_Sample1]

          sample 2 [ class B grayValue_Sample2]

          ........

          sample n [ class A grayValue_Samplen]

I hope this helps.

Dear  Jose, 

Thank very  much  for  this information. 

Do you  have  articles about that or any book title which advice  me to read.

Regards. 

Shaima 

Dear Shaima,

I wrote I book chapter that explains the use of features for image segmentation, but by using support vector machines. However, the use of the features and the vector creation is the same. It is entitled: 

• Dolz, Jose, et al. "Subcortical structures segmentation on MRI using support vector machines." Multimodal imaging towards individualized radiotherapy treatments (2014): 24.

And you will find it in my RG profile.

Additionally, there are several nice works that also show the use of other different features used in machine learning for medical image segmentation, such as:

• Powell, Stephanie, et al. "Registration and machine learning-based automated segmentation of subcortical and cerebellar brain structures." Neuroimage 39.1 (2008): 238-247.
• Magnotta, Vincent A., et al. "Measurement of Brain Structures with Artificial Neural Networks: Two-and Three-dimensional Applications 1." Radiology 211.3 (1999): 781-790.
• Kim, Eun Young, and Hans Johnson. "Multi-structure segmentation of multi-modal brain images using artificial neural networks." SPIE Medical Imaging. International Society for Optics and Photonics, 2010.
• Zhang, Nan, et al. "Multi-kernel SVM based classification for brain tumor segmentation of MRI multi-sequence." Image Processing (ICIP), 2009 16th IEEE International Conference on. IEEE, 2009.

As I told you, although deep learning and machine learning techniques are somehow different, the use of input features may remain the same.

Cheers.

Dear Jose, 

Thanks a lot,  I'm  appreciated  your help and support. My wish for you  to success  in your  research. 

Regards. 

Shaima. 

Thank you Shaima.

I hope you found useful my help.

Regards.

Thanks Jose for the detailed explanation.

So, when you said " samples are composed by (ideally) discriminative features". Do you mean features that are identified by human eyes such as edges and shapes or colours? 

In my case, I am working with histopathology images and its difficult to find discriminative features in the coarse view! does that mean Deep learning would be useful in these task?

The second question, when the actual features ( Gabor filter, or wavelet decomposition) are actully used if we are just using the grey channel  as class training sample? 

Thank you very much for this interesting discussion .