35 - Cell Nuclei analysis in Python using watershed segmentation

[@tailises1713]

Please continue this amazing job! Your tutorials are great!

[@ADPathos]

This was far more useful to me than the openCV tutorial. Thanks!

[@DigitalSreeni]

Glad to hear that!

[@adityatejaswi6178]

1:32 - 1:56 Very well said sir.

[@DigitalSreeni]

Thanks

[@felip6180]

Quite amazing seeing that the program you wrote works for other things too!

[@gye1302]

Glad to see some similar/related thing I did in my lab.

[@futurecs5471]

Hi Guochang Ye I chose to comment underneath you because it the recent post, I wish to ask that after using watershed as your segmentation, how did you do feature extraction? I would appreciate your help, even if you give me some links that helped you for feature extraction it will still mean a lot.

[@gye1302]

@@futurecs5471 Hi, for us, we were focusing on the size and distribution of the mitochondria. Not something complex features.

[@mirasalsanila]

AMAZING VIDEO THANK YOU SIR ❤❤

[@amarug]

awesome tutorial and i really like your attitude!

[@DigitalSreeni]

Glad you liked it!

[@shashikumarmaurya2930]

Nice Tutorial !!! Good Job

[@ramchandracheke]

First I like your videos then watch it !

[@DigitalSreeni]

Wow, thanks

[@Lee-vs5ez]

this helped me a lot, thanks mate

[@DigitalSreeni]

Glad it helped

[@guyhadary9372]

Hi, Great tutorial, thank you for that. It worked perfectly for 1 full folder of images and count the cells perfectly (after few adjustments). But when I tried it on different folder the counting wasn't good. Many cells with weak fluorescence weren't including in the counting, I tried to play with the threshold but couldn't change the counting result. Any suggestions could really help my project. Thank you.

[@Renardbardhi1]

Amazing job indeed. I just want to ask how can I increase the tension of the yellow color that surrounds the cells. To be more intense.

[@univweb1385]

hello sir, i am on nuclei segmentation and i need some datasets to work with, please do you have any idea howa can i get segmentation datasets such as TCGA ...

[@nying3452]

Sir, can I used this sure-foreground method in my paper to publish? Your idea is excellent. I want to cite this to avoid the academic ethic, but this is the KZbin channels. Therefore, I can't add this KZbin link in my paper according to the Journel demand.

[@DigitalSreeni]

please go ahead and use it. This is a standard well-known practice anyway :)

[@nying3452]

@@DigitalSreeni Thanks, Sir.

[@asmabenbrahem6]

Can you please answer this question: why distance transform is better than erosion ? you said that erosion will eliminate the tiny cells, but setting a big factor (a.k.a factor*max (distance_transform)) will also elminate this tiny cells. Can you please provide me with a good explanation or a link with trusted information ?

[@sciencelogs]

As much as I have tried, I have not been able to do the watershed on a single-channel grayscale image. Any solution for it. Your help would be greatly appreciated

[@krisbramleydrums2261]

Hi Sreeni, thanks so much for doing these videos, they are keeping me sane during Covid lockdown! Quick question: is there a way to automatically count the number of cells in the image following the watershed segmentation?

[@DigitalSreeni]

Thanks for your kind comments. You can use regionprops to quantify segmented regions. Please watch my video number 117 on this topic.

[@TDAnyadubaMScPhD]

This is so great! Please I have a question, I have droplets with particles inside them, how do I apply this to count the number of particles inside each droplet. Basically I need to know, area of droplet and the number of particles inside it

[@Brickkzz]

A quick work around is to create a new ROi or canvas of the droplet, then count the particles within it.

[@jamilal-idrus1905]

excuse me sir, i am always follow your another tutorial, but in this tutorial... how can i show my final watershed image in google colab?

[@BrunoGocan]

What is, if there is any, the relation between k-means cluster and watershed segmentation?

[@DigitalSreeni]

There isn't any relation. k-means is a clustering algorithm that is designed to identify a given number of clusters in your data. Watershed is an algorithm that separates different objects in an image based on the geological watershed metaphor. You can use k-means to segment objects and then watershed to separate touching objects.

[@hajobarolachattujje3831]

can you share how to do watershed segmentation of cell boundaries with the same osteosarcoma image?

[@DigitalSreeni]

This tutorial uses watershed.

[@hajobarolachattujje3831]

Yes ,but this does nuclear segmentation not cell boundaries, I was hoping if there was a way to separate the cell boundaries ?

[@pareshpriyadarshanrana2932]

Can you please tell about how to know which color in img2 corresponds to which label of "regions" while extracting information. Also, can we number label these nuclei in python the way ROIs are labelled in ImageJ?

[@DigitalSreeni]

May be my video 116 about object measurements answers your question about labeling nuclei.

[@pareshpriyadarshanrana2932]

@@DigitalSreeni The regionprops gives us label as numbers, however I need to correspond these label numbers with the color shown in img2 or "Colored grains'. For e.g., label 1 might correspond to a color of yellow in the img2. So, is there a command with which I can obtain this data.

[@hrizidorsaf2748]

please sir if you solved this problem please help me 😊

[@pareshpriyadarshanrana2932]

@@hrizidorsaf2748 well, it took a little walk around tbh, I started with one object and noted the color. Then applied the program to a 2 object image and noted the color of object "number 2". Like that added 3 objects and noted color for third one. Although I couldn't label them on the image itself, it still gives a good way of keeping track of them. Also, I must mention that the trick stops being efficient if there are more than 7 objects in a field. Ofcourse, it's not the best option out there, and hopefully you can come across some way to achieve the kind of result you want.

[@christophervrooman7580]

Cool stuff!

[@mahhhhh2599]

Hi professor. I'm doing cnn for stem cell for my final year project. But I want to know, how to train the system, since it has no memory or cannot identify which is which type of stem cell. I have CFU GM, M and G. I don't have the idea of what to do for segmentation because all the cell looks same. Can u give me some advice, on how I should process my data

[@DigitalSreeni]

If you cannot tell the difference between images then no machine learning or image processing algorithm can help. Normally, you'd manually separate images into various bins (e.g. CFU, GM, M and G) and label them accordingly. Then pick an approach to train a model. For example you can use deep learning based classification. Then use the model to classify future images. In summary, if a human cannot the difference between a cat and a dog then a machine will definitely fail.

[@mahhhhh2599]

@@DigitalSreeni thank you very much sir. Does this means like I should make a ground truth for the images, or train the class separately? I already done until the data augmentation part for all the classes. So I hope you can advice me, what kind of segmentation that I should do for CFU of the stem cells.

[@DigitalSreeni]

@@mahhhhh2599 Not sure of your exact application. Are you trying to segment pixels in an image or are you classifying whole images into various classes? For any of these you need lots of training images. Data augmentation can help a bit but yo cannot only have 100 images for each class and augment your way. You need at least 500 images per class to start seeing decent results, for classification. Once you have the data ready then you can pick any of the classification models you can copy from my github or from other accounts. The bulk of the challenge lies in getting the right data and prepping it. Actual model training is the easy part.

[@mahhhhh2599]

@@DigitalSreeni now i understand. the images that i have already in their classes. but i'm having trouble in preparing the right data., like what is the right segmentation that i should apply on these images. but i think i should do further research by reading more journals. anyway, thank you for explaining to me. thank you very much.

[@KrwaweMacki]

Hello, can I get the source of this image ? I need it for my thesis, thanks in advance

[@DigitalSreeni]

The image is on my github page: github.com/bnsreenu/python_for_microscopists Collected by my colleague on a ZEISS microscope.

[@ramyac4738]

Can this concept be applied to H&E stained images?

[@DigitalSreeni]

Yes, but you need to first separate H and E contributions. I just recorded a video on this topic so please watch the 122 video in my playlist. Direct link here: kzbin.info/www/bejne/r4bVqHiPnL-IqaM

[@ramyac4738]

@@DigitalSreeni thank you Sir! I'll try and let you know the result.

[@abdulrahimshihabuddin1119]

Ramya C I have the same question. Can this be applied onto H&E stained images??