Detecting centroid and orientation of capsules

I have a stack of 128 binary images like this one:

I want to detect centroid and orientation of capsules in 3D but, as you can see, many capsules have regions connected to other objects that I want to exclude from the detection procedure. I tried to use regionprops3, but I am not sure if this is the right MATLAB function to be used in this case. I think that, if I am able to impose the detection of my shape, I could get the expected results. Do you know if there is a method suitable for this task?

Answer 1

[Anton Semechko]

If your capsules have parametric a representation (e.g., ellipsoid), you can use Hough transform (HG). Description of a HG variant designed specifically for detecting ellipsoids can be found here: Partitioned Hough transform for ellipsoid detection

Alternatively, if you are able to segment-out individual objects from your image, you can compute for each object a set of features invariant to rigid motion (e.g., aspect-ratio, surface area, volume, etc.). You can then use these features to discriminate between capsules and other objects present in your data.

[matnewbie]

Thank you for your answer. My capsules are not ellipsoids, but can be described as a cylinder connected with two half-spherical caps. I think it is better to segment individual objects in my image but, as you can see in the image, some capsules are connected to objects (the outer black region) I want to exclude from segmentation.

[Anton Semechko]

It would help if you posted on here a sample 3D stack that you want to segment (make sure to include voxel dimensions). It is difficult to say what approach will or will not work just by looking at a single slice.

[Anton Semechko]

The sample you uploaded is a raw grayscale image that is extremely noisy compared to the relatively clean segmentation shown in your thumbnail. Segmentation of the raw image into foreground and background is the first step towards identifying individual capsules, and given the nature of the data, it will require a non-trivial solution that goes far beyond simple threholding. So, if you already figured-out how to segment the raw image data while suppressing noise, upload the segmented image. Otherwise, I suggest you revise your question.

[matnewbie]

Because in my thumbnail the image was binarized with a threshold value.

[Anton Semechko]

It seems to me that your problem is a bit more challenging than you think. For example, see if you can use thresholding to separate foreground from background in xy-slices 61 to 68.

[Anton Semechko]

To get started, I suggest you look into nonlinear filters (e.g., nonlocal means, diffusion, bilateral, median, etc.) so you can remove as much noise from the data as possible prior to segmentation.

[matnewbie]

I solved the problem by applying in 3D the procedure reported here: https://blogs.mathworks.com/steve/2013/11/19/watershed-transform-question-from-tech-support/

Then I applied regionprops3 and I got object properties.