How I can develop this segmentation codes ? I am so close to exact shape

I am using miasdatabase and this is mdb005 image segmentation process. I found this image in a scientific paper .

https://www.researchgate.net/publication/346855010_Automatic_suspicions_lesions_segmentation_based_on_variable-size_windows_in_mammography_images

And part (c) shows exact segmentation . Actually I need part c . My codes is below and result is so close to part (c) but not exactly.

What can I do to get it ? I added mdb005 image on the attachment part.

clc;    % Clear the command window.
close all;  % Close all figures (except those of imtool.)
clear;

workspace;  % Make sure the workspace panel is showing.
fontSize = 20;

imgf=imread('mdb005.png');

if(size(imgf,3)>1)
imgf=rgb2gray(imgf);
end

img=imgaussfilt(imgf);

IM=img;
subplot(3,3,1),imshow(IM);
title('Original Image', 'FontSize', fontSize);
% Maximize the figure window.
set(gcf, 'Position', get(0, 'ScreenSize'));
axis square;


thresholdValue = 135;
binaryImage = IM > thresholdValue; 
labeledImage = bwlabel(binaryImage);
measurements = regionprops(labeledImage, 'Area');
% Get all the areas
allAreas = [measurements.Area]; 
[biggestArea, indexOfBiggest] = sort(allAreas, 'descend');
% Extract biggest
biggestBlob = ismember(labeledImage, indexOfBiggest(1));
% Convert to binary
biggestBlob = biggestBlob > 0;
 subplot(3, 3, 2);
% figure,
imshow(biggestBlob, []);
axis on;
axis image; % Make sure image is not artificially stretched because of screen's aspect ratio.
title('Final', 'fontSize', fontSize);
drawnow;

maskedGrayImage = imgf; % Initialize.
maskedGrayImage(~biggestBlob) = 0;
% Display the masked grayscale image.
subplot(3, 3, 3);
imshow(maskedGrayImage);
axis on;
axis image; % Make sure image is not artificially stretched because of screen's aspect ratio.
title('Masked Gray Scale Image', 'fontSize', fontSize);
drawnow;

I_eq = adapthisteq(maskedGrayImage);
subplot(3,3,4),imshow(I_eq);
title('Equalized Image', 'FontSize', fontSize);

se = strel('disk',20);
Io = imopen(I_eq,se);
subplot(3, 3, 5);
imshow(Io)
title('Opening')

Ie = imerode(I_eq,se);
Iobr = imreconstruct(Ie,I_eq);
subplot(3, 3, 6);
imshow(Iobr)
title('Opening-by-Reconstruction')

Ioc = imclose(Io,se);
subplot(3, 3, 7);
imshow(Ioc)
title('Opening-Closing')

Iobrd = imdilate(Iobr,se);
Iobrcbr = imreconstruct(imcomplement(Iobrd),imcomplement(Iobr));
Iobrcbr = imcomplement(Iobrcbr);
subplot(3, 3, 8);
imshow(Iobrcbr)
title('Opening-Closing by Reconstruction')

fgm = imregionalmax(Iobrcbr);
subplot(3, 3, 9);
imshow(fgm)
title('Regional Maxima of Opening-Closing by Reconstruction')

se2 = strel(ones(5,5));
fgm2 = imclose(fgm,se2);
fgm3 = imerode(fgm2,se2);

fgm4 = bwareaopen(fgm3,10);
% I3 = labeloverlay(grayImage,fgm4);
figure,
subplot(2, 2, 1);
imshow(fgm4)
title('Modified Regional Maxima Superimposed on Original Image')

IM_cb = imclearborder(fgm4);
 BW2 = bwareaopen(IM_cb, 200);
BW_filled = imfill(BW2, 'holes');
 subplot(2, 2, 2);
 imshow(BW_filled);
 axis on;
axis image; % Make sure image is not artificially stretched because of screen's aspect ratio.
title('Small Lines Eliminated', 'fontSize', fontSize);

Answer 1

[yanqi liu]

sir，use the code at 《Detection-of-Breast-Cancer-using-Neural-Networks-master》

the result is

[Ali Zulfikaroglu]

this is not result

[yanqi liu]

yes，sir，use some morphological process，get the result

[Ali Zulfikaroglu]

@yanqi liu Can you write your code to get this?

[yanqi liu]

yes，sir，the code base on 《Detection-of-Breast-Cancer-using-Neural-Networks》

clc; clear all; close all;
warning off all
a=imread('demo.jpg');
b=rgb2gray(a);

%adaptive median filter part
NoisyImage=b;
NoisyImage=double(b);
[R, C, P]=size(NoisyImage);
OutImage=zeros(R,C);
Zmin=[];
Zmax=[];
Zmed=[];


for i=1:R
    
    for j=1:C
        if (i==1 && j==1)
            
            
            % for right top corner[8,7,6]
        elseif (i==1 && j==C)
            
            
            % for bottom left corner[2,3,4]
        elseif (i==R && j==1)
            
            
            % for bottom right corner[8,1,2]
            
        elseif (i==R && j==C)
            
            
            %for top edge[8,7,6,5,4]
            
        elseif (i==1)
            
            
            % for right edge[2,1,8,7,6]
            
        elseif (i==R)
            
            % // for bottom edge[8,1,2,3,4]
            
        elseif (j==C)
            
            
            %// for left edge[2,3,4,5,6]
            
        elseif (j==1)
            
            
            
            
            
        else
            
            
            
            SR1 = NoisyImage((i-1),(j-1));
            SR2 = NoisyImage((i-1),(j));
            SR3 = NoisyImage((i-1),(j+1));
            SR4 = NoisyImage((i),(j-1));
            SR5 = NoisyImage(i,j);
            SR6 = NoisyImage((i),(j+1));
            SR7 = NoisyImage((i+1),(j-1));
            SR8 = NoisyImage((i+1),(j));
            SR9 = NoisyImage((i+1),(j+1));
            TempPixel=[SR1,SR2,SR3,SR4,SR5,SR6,SR7,SR8,SR9];
            Zxy=NoisyImage(i,j);
            Zmin=min(TempPixel);
            Zmax=max(TempPixel);
            Zmed=median(TempPixel);
            A1 = Zmed - Zmin;
            A2 = Zmed - Zmax;
            
            if A1 > 0 && A2 < 0
                
                %   go to level B
                B1 = Zxy - Zmin;
                B2 = Zxy - Zmax;
                if B1 > 0 && B2 < 0
                    PreProcessedImage(i,j)= Zxy;
                else
                    PreProcessedImage(i,j)= Zmed;
                    
                end
            else
                
                if ((R > 4 && R < R-5) && (C > 4 && C < C-5))
                    
                    S1 = NoisyImage((i-1),(j-1));
                    S2 = NoisyImage((i-2),(j-2));
                    S3 = NoisyImage((i-1),(j));
                    S4 = NoisyImage((i-2),(j));
                    S5 = NoisyImage((i-1),(j+1));
                    S6 = NoisyImage((i-2),(j+2));
                    S7 = NoisyImage((i),(j-1));
                    S8 = NoisyImage((i),(j-2));
                    
                    S9 = NoisyImage(i,j);
                    S10 = NoisyImage((i),(j+1));
                    S11 = NoisyImage((i),(j+2));
                    S12 = NoisyImage((i+1),(j-1));
                    S13 = NoisyImage((i+2),(j-2));
                    S14 = NoisyImage((i+1),(j));
                    S15 = NoisyImage((i+2),(j));
                    S16 = NoisyImage((i+1),(j+1));
                    S17 = NoisyImage((i+2),(j+2));
                    TempPixel2=[S1,S2,S3,S4,S5,S6,S7,S8,S9,S10,S11,S12,S13,S14,S15,S16,S17];
                    Zmed2=median(TempPixel2);
                    PreProcessedImage(i,j)= Zmed2;
                else
                    
                    PreProcessedImage(i,j)= Zmed;
                    
                end
                
            end
            
        end
        
    end
end

imshow(PreProcessedImage,[])

%Segmentation part

Y=double(PreProcessedImage);

k=2; % k: number of regions

g=2; % g: number of GMM components

beta=1; % beta: unitary vs. pairwise

EM_iter=10; % max num of iterations

MAP_iter=10; % max num of iterations

% fprintf('Performing k-means segmentation\n');

[X,GMM,ShapeTexture]=image_kmeans(Y,k,g);
if ndims(Y) == 2
    Y = cat(3,Y,Y,Y);
end
[X,Y,GMM]=HMRF_EM(X,Y,GMM,k,g,EM_iter,MAP_iter,beta);

Y=Y*80;

Y=uint8(Y);
figure, imshow(Y);

bw = im2bw(rgb2gray(Y));
bw = imclearborder(bw);
bw2 = imopen(bw, strel('line', round(size(bw,2)*0.05), 0));
bw3 = bwareafilt(bw2, 1);
be1 = PreProcessedImage;be2 = PreProcessedImage;be3 = PreProcessedImage;
be1(bw3) = 255;be2(bw3) = 0;be3(bw3) = 0;
be = cat(3,be1,be2,be3);
figure; montage({bw3,mat2gray(be)}, 'Size', [1 2], 'BackgroundColor', 'w', 'BorderSize', [2 2]);

[Ali Zulfikaroglu]

It doesn't work. Because tag and pectoral muscle should be removed some way and then you should find tumor suspicious part

[yanqi liu]

yes，sir，because the image exist white block in upper right，may be use the image

https://www.mathworks.com/matlabcentral/answers/uploaded_files/812659/image.jpeg

to test

or upload your image to do some analysis

[Ali Zulfikaroglu]

@yanqi liu thank you for your struggle. But mammography images comes with usually label and pectoral muscle. That's why I have to segment with them

Answer 2

[Image Analyst]

I didn't read the paper but assuming you implemented the functions correctly, the problem may be that your image sizes don't match theirs and you need to adjust some of the image processing parameters like window sizes or min acceptable blob size or something like that.