How to segment image without losing information at the edges for finding centroid?

I have the following image and when I segment and fill it, I'm getting some error which ends up in the image having mutiple centroids.

How to I fix it make sure that I get only one centroid for the entire image. I have attached the code of my progress till now

clc;clear;close all;

I=imread("Abdomen CT.jpeg");I=im2gray(I);
imshow(I); title('Original Image')

level = graythresh(I);
BW = imbinarize(I,'adaptive','Sensitivity',level,"ForegroundPolarity","dark");
imshow(BW,[]); title('Segmented Image')

BW2 = imfill(BW,'holes');
figure
imshow(BW2)
title('Filled Image')

Trace region boundaries in binary image

[B,L] = bwboundaries(BW,'noholes');
imshow(label2rgb(L, @jet, [.5 .5 .5]))
hold on
for k = 1:length(B)
   boundary = B{k};
   plot(boundary(:,2), boundary(:,1), 'w', 'LineWidth', 2)
end

Centroid calculation

s = regionprops(BW,'centroid');
centroids = cat(1,s.Centroid);
plot(centroids(:,1),centroids(:,2),'b*')

Answer 1

[Bjorn Gustavsson]

The only way I can imagine the meaning of centroid in a grayscale image like the one above is with the same definition as "centre of mass" has. For that you could do something like this:

[X,Y] = meshgrid(1:size(I,2),1:size(I,1));
Id = double(I)
r_centroid = [sum(X(:).*Id(:)),sum(Y(:).*Id(:))]/sum(Id(:));

This would properly give you a centre-of-brightness of the abdominal slice, which is not dependent on any arbitrary thresholding:

HTH

[Image Analyst]

This is definitely the way to go.

This is how I'd do it, because you can't simply threshold and get centroids of all the individual blobs and average those or get the convex hull and get the centroid of that.

Your image is extremely low resolution. You should use a higher resolution image, then identify the body and table and mask out the table and surrounding air, then use Bjorn's method.

I = imread('image.png');
if ndims(I) == 3
    I = rgb2gray(I);
end
imshow(I, []);
axis('on', 'image')
% Crop off huge white frame
if I(1,1) == 255
    frameRows = all(I == 255, 2);
    I(frameRows, :) = [];
    frameColumns = all(I == 255, 1);
    I(:, frameColumns) = [];
end
imshow(I, []);
% Compute centroid.
[X,Y] = meshgrid(1:size(I,2),1:size(I,1));
Id = double(I);
r_centroid = [sum(X(:).*Id(:)), sum(Y(:).*Id(:))]/sum(Id(:))
hold on;
plot(r_centroid(1), r_centroid(2), 'r+', 'LineWidth',3, 'MarkerSize', 150)

[Max V]

@Bjorn Gustavsson @Image Analyst Thank you for your suggestions.

@Image Analyst,

Actually the problem statement given to me is:

Employ methods of image segmentation, texture analysis, etc. in order to identify the type of medical image. You cannot employ any machine learning / deep learning method in order to achieve a solution. The files should be submitted as independent to run functions in MATLAB.

The dataset provided (https://www.kaggle.com/andrewmvd/medical-mnist) has only images of the resolution 64x64px.

I visualized appraching this problem by as follows:

• Segment the image (and obtain proper boundaries)
• Find the centroid of the Image.
• Find the distance of the edges from the centroid for 0-360 degrees.
• Based on these values create a threshold and use conditional statements to identify and classify the image types.
• I also intended to use GCLM values and one of your previous code shape_recognition_demo2.m to extract perimeter values and use it as additional checks.

I'm not sure how efficient this approach is as I'm just starting out with MATLAB and coding in general.

Any suggestions from your end would be welcome to improve the current approach or to try out any new ones.

Thank you for your time and help in advance!

[Bjorn Gustavsson]

@Max V: "texture analysis" - the way I've understood "texture" is that it is the small-scale (quasi-periodic) pattern of a surface, for example typical differences between cloth of different textiles (tweed, denim, velvet etc), or of satellite-images of different surfaces (grass-land, forests, sea with different types of waves, or who knows how many strange types of ice). In this image there cannot possibly be any type of sensible "textures", surely? What I guess is visible in your image is a cross-section of a chest, with ribs and spine visible and then a heart? Those aren't what I'd call textures. You could extract the boundaries of different whiter blobs either by contour or imbinarize. However, to "in order to identify the type of medical image" from images of this resolution, as pointed out to be very low by @Image Analyst, you'll have to settle for lower ambitions, perhaps settling for noting that this is a CT-image of an abdomen.

[Image Analyst]

To get edges you can use edge() or imgradient(). To find the coordinates of the edges you can use find(). Then find the gray levels from the image. The thing is you'll need to know if the edge is on the dark side or bright side of the edge because that could make a huge difference on the threshold you pick. For example if you had a solid uniform bright disk and you looked inside the disk, you might get a bright threshold that, when used, would not segment the bright disk. So you need to use the darker value near the edge if you want to threshold to get bright things. You could use imerode() to do a local min and then read the gray level of the edge from the eroded image.

To get the distance from the centroid all the way around, you can pass the edge x and y coordinates into sqrt().

distances = sqrt((x-xCentroid).^2 + (y - yCentroid).^2);

[Max V]

@Image Analyst thank you for your response!

If I am to classify these images based on Texture analysis will the GCLM properties like Contrast, Correlation, Energy, Homogenity etc suffice or shall I also use LBP (Local Binary Pattern) and Law's Texture Energy Measures too for accuracy?

[Bjorn Gustavsson]

@Max V: In the example image you've presented the only "Local Binary Pattern"s I can see are the regions with 2-pixel-wide interference-like patterns (either from the CT-operation, too sparse number of projections or too high noise-level), or from a horrendously rough jpeg-compression. Here you'll have to settle for size, orientation, shape and widths of the different regions, you could perhaps take brightness-profiles of the different regions at the narrowest point.

Answer 2

[yanqi liu]

sir,may be use imclose to one block area

clc;clear all;close all;

I=imread("https://ww2.mathworks.cn/matlabcentral/answers/uploaded_files/802049/Abdomen%20CT.jpeg");
if ndims(I)==3
I=rgb2gray(I);
end
imshow(I); title('Original Image')

level = graythresh(I);
BW = imbinarize(I,'adaptive','Sensitivity',level,"ForegroundPolarity","dark");
imshow(BW,[]); title('Segmented Image')

BW2 = imfill(BW,'holes');
figure
imshow(BW2); hold on;
title('Filled Image')
s = regionprops(BW2,'centroid');
centroids = cat(1,s.Centroid);
plot(centroids(:,1),centroids(:,2),'b*')

% close bw
BW3 = imclose(BW2, strel('disk', round(mean(size(I)/2))));
figure
imshow(BW3); hold on;
title('Closed Image')
s = regionprops(BW3,'centroid');
centroids3 = cat(1,s.Centroid);
plot(centroids3(:,1),centroids3(:,2),'r*')

% distance
dis = centroids - repmat(centroids3(1,:), size(centroids,1), 1);
dis = dis(:,1).^2 + dis(:,2).^2;
[~,ind] = min(dis);

% display
figure;
imshow(BW2); hold on;
title('Filled Image')
s = regionprops(BW2,'centroid');
plot(centroids(ind,1),centroids(ind,2),'r*')

[Max V]

Thank you for taking time to reply @yanqi liu!

I want to also perform Shape Signature analysis by calculating the distance and angle from the centroid to the edge of the image.

Can you help/guide me on how to do it?