Edge segmentation of cell images

segmentation

#1

Hello, I am a newcomer to cell profiler. My job is to do deep learning image processing. I have a problem now. I need to label cell images and then use deep learning to train, but some cell images are closely distributed. I can’t label them one by one, so my idea is that before I mark, I split the image and then mark it. I tried to split by skimage, but the effect is not very good. Does the cell profiler have corresponding pipelines that can directly segment the cell edge on the original image??

My original picture is like this:

I want to automatically segment the edge between each cell on the original image. Can the cell profiler do it? If you can please tell me, thank you. If not, is there any other code method that can be done?


#2

Not related to segmentation:
May I ask how many images you have for the training procedure?

Related to segmentation:
Are you interested in the small and bright structures or in the large and faint ones?

Regards

Herbie


#3

CellProfiler probably can do it, though not with the 100% accuracy you’re hoping for deep learning- since most of its algorithms rely on objects having bright centers, I’d try inverting the image and segmenting on that. If you need to do some cleanup, you can then use the EditObjectsManually module.


#4

Hello, thank you for your reply, but I think you may not understand what I mean.
My goal is to automatically segment the edges and areas of the image. Just like a map, every country has its own outline. I want to automatically split the edges between the cell membranes, just like a map, then I Then mark the outline of this image, this has nothing to do with deep learning. My ultimate goal is to use the image processed in this way for deep learning.
Can this description be understood?


#5

Oh, I well understood your request and my reply has two headings…

You didn’t reply to none of my two questions though.

However, indirectly I conclude that you are interested in the segmentation of all structures (bright and faint), which won’t be easy.

The provided sample image has a strange discrete histogram. It appears suboptimum for most kinds of evaluation and at least requires some explanations about its acquisition process.
Is the sample image perhaps taken from the literature? If no, how did you acquire it?

I’m still interested in an answer to my less related question #1.

Regards

Herbie

PS:
Here is the best what I could get from your sample image by some pre-processing:


This pre-processed image is not sufficient to get what you want to see in the end. Noise may be further reduced but the main problem is that in the original image most cell membranes are at least partly much too faint to get closed cell contours.

Your image suffers from a rather small dynamic range of the gray values and from a spatial resolution that is at the lower limit for the task in question.


#6

Hello, my sample data comes from Human cells in your celeprofiler website. This is similar to the real image I need. I saw the sample image you provided, which is close to the result I want, if I Can get high resolution images, I believe will get very clear results
Is it convenient for you to share your method or code case with me? Thank you!


#7

I’m looking forward to seeing the actual images that you want to analyze.
If you post representative sample images in the original TIF- or PNG-format we shall see which kind of processing is promising and if reasonable cell-contours can be extracted.

[…] I believe will get very clear results […]

I prefer not to believe in something I haven’t scientificly investigated …

Furthermore, I’m still curious concerning my first question.

Regards

Herbie

PS:

my sample data comes from Human cells in your celeprofiler website

That’s interesting and I really would like to know if CellProfiler is able to do the segmentation, i.e. to provide closed cell contours for all of the cells.


#8

Regarding your first title: related to segmentation, I have no training data yet.
In addition, the actual image I want to split is as follows:
%E4%B8%8B%E8%BD%BD%20(2)

If I look at the naked eye and then mark each cell in the image, it is very difficult, so I want to first segment the outline between the cells, then mark it, can you share your method with me? Thank you


#9

Both images have nothing in common and I see no way to present an single approach that treats them with the desired result.

Furthermore, the images are now in color and both show a completely different staining. In the first one I see a chance to detect and isolate the nuclei but not the cell boundaries. The second one shows nearly the same deficiencies as the achromatic first one.

BTW, using JPG-compressed images is a no-no for scientific investigations.

Are these really the images you have acquired yourself?

Regarding your first title: related to segmentation, I have no training data yet.

The meaning of my first question was: How many images do you expect to need for your training procedure?

I think that this is an essential question because you must acquire a rather large number of images and presently I don’t see how this could happen.

Good luck

Herbie


#10

I expect the amount of data to be trained is not my concern, this is the data I found online, the real picture format is tiff, because I did not find the cell image of the cell membrane, so use these two pictures instead, but the real cell membrane and The images I provided at one time are similar, so I want to know how the split image you provided above was obtained.


#11

I expect the amount of data to be trained is not my concern

That’s funny, because, as I understand, you need to provide the segmented images — no?

this is the data I found online,

I asked you to wait with the processing until you have your own data.

the real picture format is tiff, because I did not find the cell image of the cell membrane, so use these two pictures instead

I don’t understand you here.

I want to know how the split image you provided above was obtained.

My method won’t work with these images and I’ve explained why this is so.

Please understand that I won’t engage any further in this topic until you present your own images in a suitable file format and with a suitable dynamic range and spatial resolution.

Regards

Herbie


#12

Since you said you wanted to do this with CellProfiler, essentially the approach I would recommend for your second and third images is (if you want to do a segmentation)

  1. IdentifyPrimaryObjects to find your Nuclei
  2. IdentifySecondaryObjects to find your Cells
  3. ConvertObjectsToImage
  4. SaveImages

If you literally just want enhance the edges, you can do so with the EnhanceEdges module.

Again, though, as I said before, depending on your use case (is the ultimate objective you need these for an object detection problem? a segmentation problem?) you almost certainly need more accurate boundaries than any method can auto-generate, which is why you’ll need to either a) Get an approximate segmentation and tweak it with something like EditObjectsManually or b) Generate entirely manual annotations, with something like IdentifyObjectsManually or GIMP.


#13

Hello, sorry, I have seen it for so long.
I really want to do the segmentation. I want to split the cell image by cell profilter, but I don’t know how to do it. Here is a real cell fluorescence image. Take this image as an example. How should I do it through the cell profiler? Split?
The image is as follows:
img_1881226


#14

Hello, I used the pipeline method you provided to segment the cell image, but I found that many cells were divided into several pieces. What I want is a complete segmentation of the cells. Do you have any better suggestions? ?
Here are the results of my method of testing your pipeline with the human body cells provided by the cell profiler: