Chan Vese Segmentation mask is not working Properly

I cropped my DICOM image so the active contour be faster and segment between the black area and soft tissue.

Instead, I got a contour that overlaps the edge of the image and goes beyond this. How can this happen? See attached image.

Let me know if you need the m files.

Answer 1

[Anton Semechko]

Your image appears to be "padded" on the top and bottom with gray values. The transition between the black background (containing the region of interest; ROI) and these gray regions, has large gradient that will attract the contour during segmentation; because evolution of the contour is guided by the strength of the image intensity gradient. My suggestions to you, (1) remove the padding, (2) initialize contour close to the ROI you want to segment.

[Stelios Fanourakis]

You are right. But how do I do that? I cropped the image without padding still the padding shows up and I am trying to find a way to initialize the contour close to the dark area. If you have any experience on chan vese methods please help

[Image Analyst]

"I cropped the image without padding still the padding shows up" Huh? How could that be? Attach your m-file and dicom image, zipped up if you have to. And you've Accepted this answer, so can we assume that it's solved, or not? Please clarify.

[Stelios Fanourakis]

I accepted the answer because he's right to his comment, I guess. No it hasn't been solved. I attach the two m files I use for the contour and the jpg (cropped) image. You'll see that the contour goes outside the image where I want it to go inside of it and match the black area. It must be a tune in the init_mask of the chvese.m but I cannot find it yet.

[Image Analyst]

I don't see any cropping going on there.

[Stelios Fanourakis]

The image is already cropped. The original one is much bigger. I only left a thin strip of the dicom image and saved it as jpg

[Stelios Fanourakis]

This white padding thing do not exist in the cropped image when opened using imtool

[Stelios Fanourakis]

I saved it as png and now no white padding is showing and the contour goes inside the image. But it seems it cannot segment between the soft tissue (gray) and the background. It outlines the whole image

[Anton Semechko]

1) Can you post on here the grayscale image you want to process (without the padding) and a separate image showing (approximately) the ROI you want to segment.

2) Is there a reason you chose this particular segmentation algorithm?

[Stelios Fanourakis]

@Anton. I just saved it as png (same image). I need an active contour to separate black area (bckgnd) with the soft tissue (foreground). It seems fast enough this algorithm. Do you have any alternative to propose me?

[Anton Semechko]

The image background appears relatively uniform and free of noise, so yes, there are simpler and more efficient algorithms that can be used in place of level-set segmentation.

Before we get to that, it seems that you are not getting an important point about the input image you want to process. The image you provided is definitely not an original DICOM image because it contains extraneous padding around the borders:

You should not be using this image as input to a segmentation algorithm, level-set or otherwise, because it contains information that will "confuse" segmentation algorithm and preclude it from finding the ROI you actually want to segment. Do you understand this?

[Stelios Fanourakis]

Yes Anton. I understand it. If you see my latest attachment it is the .png format with no padding at all. Still cannot differentiate between black and gray.

[Stelios Fanourakis]

Can you please give me using a more efficient segmentation algorithm?

[Anton Semechko]

'untitled.png' WAS the image I used to generate the above thumbnail showing the extraneous padding!!!

[Stelios Fanourakis]

But I use that image to my pc/Matlab and the contour goes inside the image area. Correctly. Previously with the 'pictoedit.jpg', there was a problem with padding. Are you sure you used the png version? I don't have problem with padding. I have problem that it won't segment correctly

[Anton Semechko]

I can't figure out if I am being trolled. Anyone else reading this care to comment?

[Stelios Fanourakis]

See my previous attachment. Contour get inside the image. How else to explain it?

[Anton Semechko]

Ok, do this:

 im=imread('https://www.mathworks.com/matlabcentral/answers/uploaded_files/122991/untitled.png');
 figure('color','w')
 imshow(im)

What do you see?

[Stelios Fanourakis]

At my first attachment in this post (fig.jpg) contour goes out of image area. That was my problem now I have a different problem. Am I clear or not?

[Anton Semechko]

Stelios, does the file 'untitled.png' contain the image you are using as input to the segmentation algorithm?

[Stelios Fanourakis]

Yes, I use the 'untitled.png', now

[Stelios Fanourakis]

@Anton. Please, give me an efficient segmentation algorithm. The one I use doesn't seem to be able to differentiate between black and gray values

[Anton Semechko]

Since you are completely impervious to everything I said up 'till now, I will just give you a demo.

% Get the image
im=imread('https://www.mathworks.com/matlabcentral/answers/uploaded_files/122991/untitled.png');

% Remove extraneous padding and border drawn on the image
im=im(161:280,11:880,:);
im=max(im,[],3);

% Segment into 4 classes using only intensity information
[C,U,LUT,H]=FastFCMeans(im,4);
L=LUT2label(im,LUT);

% Visualize
figure
subtightplot(3,1,1), imshow(im)
subtightplot(3,1,2), imshow(L,[])
subtightplot(3,1,3), imshow(L>1,[])

Segmentation functions used in this demo can be downloaded from here

'subtightplot' function can be found here

[Stelios Fanourakis]

@Anton. Although I got all the functions you suggested me, here are the results. I didn't manage to get your results. I did only copy paste.

[Anton Semechko]

Sorry about that. I used slightly modified functions. You can get them here.

[Stelios Fanourakis]

Thanks a lot Anton, I do really appreciate your effort and help. One last thing I might need to ask is how can I have the contour visible on the image. I need to control the active contour so will be slightly modified when applied to several images since the images differ between them

[Anton Semechko]

'FastFCMeans' performs image segmentation by clustering of image intensities and is thus fundamentally different from the level-set segmentation algorithm you were using before. There is no initialization involved and the resulting segmentation depends only on the input parameters 'c' and 'q'; 'c' controls the number of intensity clusters and 'q' is a fuzzification parameter. See 'FastFCMeans' documentation for additional info.

[Stelios Fanourakis]

Hi Anton.

Thanks for your valuable help so far. What I have not understand yet is how to get what I need. FastFCMeans algorithm indeed segmented the gray soft tissue from the background. What I want now is to get the black area above the soft tissue. Only this region above the tissue. I need to somehow segment it and measure its length, pixel area size, dimensions. Do I have to implement a snake contour like chan vese. How else can I obtain the values of this area? (pixel size, pixels in number, dimensions, length from top of the image to the beginning of the tissue).

[Stelios Fanourakis]

Moreover, if you notice the soft tissue at the last pic available (untitled.png) you'll notice some black gaps between the soft tissue especially at the lower left side of the image. I need some kind of interpolation method to fill those gaps so the contour will not get messed. So the soft tissue will be a unit of seamless gray or white matter with no black gaps.

I attach an image to see the results so far. See that the contour gets troubled to the points where the soft tissue is not seamless and black gaps occur. Any idea of how to fix it?