What should I do if the brightness of backgroud is very close to foreground?

Hi,

I have some images that I want to segment. However, The the brightness of backgroud is very close to foreground in my images.

Here is a example,

The red circled object is what I want. But after trying different methods, this is what I got

Is there any suggestions to get a better result? These segmented images are used to be mask for 3D segmentation by using ‘activecontour’. Other images have similar problem. And as you can imagine, The 3D structure is much larger than what I want.

Thanks!

Answer 1

[Image Analyst]

They don't look that close. Thresholding should work fairly well. Otherwise use SegNet deep learing if you need to identify triangular protuberances. Not sure what you want though. What you got looks OK to me. Why is it not? What do you need to measure from it. Tell me so I can know how important holes and shape are.

[Dongyu Liang]

Hi,

Thank you for the reply.

The only foreground area is what you mentioned, the triangular protuberances. The area out side the red circle should be considered as background. And based on that binary image mask, ‘activecontour’ can not give correct shape. what I want to measure is the internal and external surface area, solid and holes volume.

I was thinking about deep learing before, but I did not find example how to build training, validation, and test groups. Like I need to manually label different area for hundreds of images?

Thanks!

[Image Analyst]

I'd call imfill() then call bwboundaries(). Then pass the leftmost and rightmost coordinates (say the outer 10 columns or whatever) into polyfit() to get a parabola. Then use poly2mask() to get a mask of the parabola corner blob and use it to erase the original blob leaving only the triangle. Something like (untested)

mask = imfill(mask, 'holes);
mask = bwareafilt(mask, 1); % Take largest blob only.
% Get boundary
boundaries = bwboundaries(mask);
boundaries = boundaries{1};
x = boundaries(:, 2);
y = boundaries(:, 1);
% Determine leftmost and rightmost coordinates.
[rows, columns, numColors] = size(mask);
indexes = x < 10 | x > columns - 10;
% Fit a parabola
coefficients = polyfit(x(indexes), y(indexes), 2)
xFit = 1 : columns;
yFit = polyval(coefficients, xFit);
% Add bottom line
xFit = [1, xFit, columns, 1];
yFit = [rows, yFit, rows, rows];
parabolaMask = poly2mask(xFit, yFit, rows, columns);
% Get mask without parabola
mask(parabolaMask) = false;
imshow(mask);

[Dongyu Liang]

Hi,

Thank you for the helpful idea. I will try working on it.

May I ask how to create groups for deep learning?

https://www.mathworks.com/help/images/segment-3d-brain-tumor-using-deep-learning.html

This tutorial only mentions to download exist data. What should I do if I want to use deep learning in my case?

Thanks!

[Image Analyst]

I can't really explain it here. You're just going to have to go through some of the demos yourself and adapt them.

[Dongyu Liang]

thanks!