How to calculate random shift intensity difference (RSID) feature?

Hi,

I have 3D patients volumes saved in a matlab variable.

I need to extract random shift intensity difference (RSID) feature, this feature compares the intensity of the current voxel x and another

voxel x + u with random offset u, ` f(x, u) = I(x + u) − I(x)`, where u is a random offset vector.

My question is

1. How many neighbors of a specific voxel needs to be considered to calculate RSID?
2. How to find them with matlab code?

Your help is appreciated

Answer 1

[Image Analyst]

I think you'd have to scan your 3-D value one voxel at a time.

If you just want the intensity difference over all the neighbors, you can simply use convn with a kernel of all -1 excel the center being 26

kernel = -1 * ones(3,3,3);
kernel(2,2,2) = 26;
kernel = kernel / 26; % So output is in same intensity range as input.
output = convn(inputImage, kernel, 'same');

Why do you want the difference at just one randomly located pixel instead of the average over ALL neighbors?

[Sara Salimi]

Thanks a lot for your help,

What about u here (for example if u=2)?

[Image Analyst]

I guess you don't want to answer my question, but want me to answer yours.

You can get u from randi():

u = rand(uMax, 1, 1);

After that, it's just a dumb, brute force triple nested for loop that I'm sure you can do. Get a u for each direction: row, column, and slice. So you'll have 3 u's. And make sure you get the indexing right: arrays are not indexed array(x,y,z), they're indexed array(y,x,z).

[Sara Salimi]

Hi again Sir,

I apologize because I missed out the last line of your question "Why do you want the difference at just one randomly locatlcued pixel instead of the average over ALL neighbors?"

I have calculated 138 different features, I saw in one paper that authors in this paper , mentioning about semantic feature (random shift intensity difference) and semantic context feature ( random shift probability difference (RSPD), which is the probability maps from a previous stage of voxel classification ). Which I do not understand how they are calculating for each voxel with 26 neighbors.

[Image Analyst]

So let's see your for loop. Did it look something like this:

for slice = 1 : slices
    for col = 1 : columns
        for row = 1 : rows
            ur = 2*randi([0, 1], 1, 1) - 1;
            uc = 2*randi([0, 1], 1, 1) - 1;
            us = 2*randi([0, 1], 1, 1) - 1;
            randomPixel = double(image3d(row+ur, col+uc, slice+us));
            thisPixel   = double(image3d(row,    col,    slice))
            diffmage(row, col, slice) = thisPixel - randomPixel;
        end
    end
end

If not, what did it look like?

[Sara Salimi]

Could I ask is image3d the input 3D array of image, and diffimage is the output? Once I am running the last code, I am getting an error:

Subscript indices must either be real positive integers or logicals.
Error in  (line 10)
            randomPixel = double(inputDataVol(row+ur, col+uc, slice+us));

The intensity range for the original image has been normalized between [0,1].

Once I am running the first code, After applying the following code,

kernel = -1 * ones(3,3,3);
kernel(2,2,2) = 26;
kernel = kernel / 26; % So output is in same intensity range as input.
output = convn(inputImage, kernel, 'same');

I get an image like for RSID that I have show for a specific slide and the ranges of values is between [-0.3517:0.6728]

and for RSPD I get the following image for a specific probability map of structure, the values ranging [-0.6063: 0.6893]

[Sara Salimi]

Thanks for your help,

I found an easier way to do this:

   shiftRange=[-1,1];    % defining a shift range 
    u=[randi(shiftRange),randi(shiftRange),randi(shiftRange)];
    I2 = imtranslate(I,u);
    Id=imabsdiff(I,I2);  % the difference between original image and shifted imag
    

However I a not sure should I shift with a bigger range or not, sth like this:

    u=2*[randi(shiftRange),randi(shiftRange),randi(shiftRange)]-1;
    

makes the shift range changes between [-3,3]