Count and measure yeast cells from an image

Hello. I am trying to detect, count and measure yeast cells from various images like in this example,

First I made the background uniform using a morphological opening, followed by an increase of the contrast using "imadjust". Then, I found the edges of the image using the 'canny' function to create a mask that overlays on the adjusted image to remove all the background and some of the cells glow. Next, I turned the pixels below 50 into black and converted the image into a binary one. Finally, I cleaned the image and used the “watershed” transformation to separate the connected cells.

Here’s my code:

        clear all
        close all

        se1=strel('disk',2);

        X=imread('t132-1.png');
        I=X(:,:,2);
        figure
        imshow(I)
        title('original')
        drawnow

        %%background removal
        background = imopen(I,strel('disk',15));
        I2=I-background;
        I3 = imadjust(I2);

        % figure
        % imshow(I3);
        % title('Adjusted image')
        % drawnow

        %%edge detection
        Iedge=edge(I3,'canny',0.3);
        Iedge=bwmorph(Iedge,'dilate');
        Iedge=bwmorph(Iedge,'bridge');
        Iedge=imfill(Iedge,'holes');

        % figure 
        % imshow(Iedge)
        % title('edge')
        % drawnow

        Iclean=bwmorph(Iedge,'erode',2);
        Iclean=bwareaopen(Iclean,500);

        % figure
        % imshow(Iclean);
        % drawnow

        %%Image masking
        Icut=255-I3;
        Icut(Iclean==0)=0;
        figure;
        imshow(Icut);
        title('cut')
        drawnow

        Iadjust=Icut;
        Iadjust(Iadjust<50)=0;
        % figure
        % imshow(Iadjust);
        % title('black pixels below 50');
        % drawnow

        Ibw=im2bw(Iadjust,0.1);
        % figure
        % imshow(Ibw);
        % title('BW');
        % drawnow

        Ibw=imerode(Ibw,se1);
        Ibw=imdilate(Ibw,se1);
        Iclean2=bwareaopen(Ibw,150);
        % figure
        % imshow(Iclean2);
        % title('clean')
        % drawnow

        %%watershed function
        D=-bwdist(~Iclean2);
        mask=imextendedmin(D,7);
        D2=imimposemin(D,mask);
        Ld2=watershed(D2);
        bw3=Iclean2;
        bw3(Ld2==0)=0;
        figure
        imshow(bw3)
        title('Final')
        drawnow

I plan to make the measurements using the “regionprops” command. My problem is that the resulting image (shown below) lost many of the cells forms and have outter holes that difficult the measuring process. I’m wondering if you have any suggestion for a better general detection process that I could apply on all my 36 images, thanks in advance!

Answer 1

[Image Analyst]

Maybe try either imfill() or bwconvhull().

[Luis Farías]

Thank you for the suggestion, but unfortunately I’m still not getting the desired results. Let me explain. I tried both methods. First, with imfill(), not all the cells are filled because some of the inner holes aren’t closed and also the cells edges still appear “nibbled”. On the other hand, with the bwconvhull() command I get a better looking result, with rounded shapes, but many of the cells merge into bigger objects, so I cannot measure and count them individually.

[Image Analyst]

Maybe you can try marker controlled watershed segmentation, like in Steve's blog: http://www.mathworks.com/help/images/examples/marker-controlled-watershed-segmentation.html

[Luis Farías]

Thanks, I followed the Steve’s tutorial and I managed to improve the final image using the reconstruction by opening and closing thresholding. I was wondering if you think in a better method to remove the background and the cell’s glow in the first part of my code (from I ‘original’ to Icut ‘cut’ images), because I am getting an image with high contrast (making the cells full of holes) and some missing borders.

[Image Analyst]

Did you try imfill() at a strategic place in your algorithm? I might be able to improve on it, but it would take a long time. Sorry but it would take more development time than I can spend on it. I'd just try imfill() at some point and vary the parameters in Steve's algorithm to see if you can improve it. Otherwise if you just want a count you can use manual counting with ginput() or impoints() or else just go with area fraction as your thing to measure.

[Luis Farías]

Ok, I'll try that. Thanks for everything.