{"id":60,"date":"2006-06-02T06:54:41","date_gmt":"2006-06-02T10:54:41","guid":{"rendered":"https:\/\/blogs.mathworks.com\/steve\/?p=60"},"modified":"2019-10-22T12:26:39","modified_gmt":"2019-10-22T16:26:39","slug":"cell-segmentation","status":"publish","type":"post","link":"https:\/\/blogs.mathworks.com\/steve\/2006\/06\/02\/cell-segmentation\/","title":{"rendered":"Cell segmentation"},"content":{"rendered":"
\r\n

Blog reader Ramiro Massol asked for advice on segmenting his cell images, so I gave it a try. I'm not a microscopy expert,\r\n though, and I invite readers who have better suggestions than mine to add your comments below.\r\n <\/p>\r\n

Let's take a look first to see what we have. I'm going to work with a cropped version of the original<\/a> so that the images aren't too big for the layout of this blog.\r\n <\/p>\r\n

Note: you can download the functions imcredit and imoverlay<\/a> from MATLAB Central.<\/i><\/p>

I = imread('https:\/\/blogs.mathworks.com\/images\/steve\/60\/nuclei.png'<\/span>);\r\nI_cropped = I(400:900, 465:965);\r\nimshow(I_cropped)\r\nimcredit('Image courtesy of Dr. Ramiro Massol'<\/span>)<\/pre> 
Strictly speaking, contrast adjustment isn't usually necessary for segmentation, but it can help the algorithm developer see\r\n      and understand the image data better.  This is a fairly low-contrast image, so I thought it might help.  You can adjust the\r\n      display contrast interactively with imtool, or you can use an automatic method such as adapthisteq. adapthisteq<\/tt> implements a technique called contrast-limited adaptive histogram equalization<\/i>, or CLAHE.  (I always thought  \"CLAHE\" sounded like it must be some Klingon delicacy.)\r\n   <\/p>
I_eq = adapthisteq(I_cropped);\r\nimshow(I_eq)<\/pre> 
So what happens if we just apply a threshold now?<\/p>
bw = im2bw(I_eq, graythresh(I_eq));\r\nimshow(bw)<\/pre> 
Let's clean that up and then overlay the perimeter on the original image.<\/p>
bw2 = imfill(bw,'holes'<\/span>);\r\nbw3 = imopen(bw2, ones(5,5));\r\nbw4 = bwareaopen(bw3, 40);\r\nbw4_perim = bwperim(bw4);\r\noverlay1 = imoverlay(I_eq, bw4_perim, [.3 1 .3]);\r\nimshow(overlay1)<\/pre> 
Now, I'm not familiar with these cell images, so I don't know exactly what I'm looking at.  I assume some of these blobs need\r\n      more help to be separated properly.  One possible approach is called marker-based watershed segmentation<\/i>.  There's a demo<\/i> of this idea on The MathWorks web site.\r\n   <\/p>\r\n   
With this method, you have to find a way to \"mark\" at least a partial group of connected pixels inside each object to be segmented.\r\n      You also have to mark the background.\r\n   <\/p>\r\n   
Let's try to use the bright objects, which I assume are nuclei.  The extended maxima<\/i> operator can be used to identify groups of pixels that are significantly higher than their immediate surrounding.\r\n   <\/p>
mask_em = imextendedmax(I_eq, 30);\r\nimshow(mask_em)<\/pre> 
Let's clean that up and then overlay it.<\/p>
mask_em = imclose(mask_em, ones(5,5));\r\nmask_em = imfill(mask_em, 'holes'<\/span>);\r\nmask_em = bwareaopen(mask_em, 40);\r\noverlay2 = imoverlay(I_eq, bw4_perim | mask_em, [.3 1 .3]);\r\nimshow(overlay2)<\/pre> 
Next step: complement the image so that the peaks become valleys.  We do this because we are about to apply the watershed\r\n      transform, which identifies low points, not high points.\r\n   <\/p>
I_eq_c = imcomplement(I_eq);<\/pre>
Next: modify the image so that the background pixels and the extended maxima pixels are forced to be the only local minima\r\n      in the image.\r\n   <\/p>
I_mod = imimposemin(I_eq_c, ~bw4 | mask_em);<\/pre>
Now compute the watershed transform.<\/p>
L = watershed(I_mod);\r\nimshow(label2rgb(L))<\/pre> 
I don't know if this is a good segmentation result or not, but I hope some of the methods I've shown will give Dr. Massol\r\n      some ideas to try.\r\n   <\/p>\r\n   
Other readers ... if you have suggestions, I invite you to post your comments here.<\/p>\r\n