How to automatically separate fibers from nuclei in a histological image ?

Hi everybody !

I am working on histological images on matlab, and I would like to measure orientation of fibers on my images.

I have tried my orientation determination code on a 'simplistic' image from paint (see bellow) and it's working quite well. I am using structural elements lines that I rotate from 1 to 180° using imopen.

So know my next step is to isolate fibers in my pictures so that I can run the algorithm without measuring nucleus orientation for example.

Here is an example of the kind of images I have :

(fibers I would like to measure are the small black lines)

and without visible fibers :

For know, I have been struggling with this problem using various of thecnics such as opening with small disks, sobel filter with bwareaopen, or even clustering in L*a*b space for the image with colours.

I would like to have your point of view on how I could do such thing (maybe neural network is my only option here?). At first I would like to have a simple algorithm, even if it's no perfect !!

Thank you very much in advance for your help !

Simplistic image :

Answer 1

[Maneet Kaur Bagga]

Hi Francois,

As per my understanding of the question, to isolate the fibres from the histological images considering a neural netwrok based approach can be a suggested option. For designing a neural network for the fibre segmentation you may use popular architectures like U-Net, FCN(Fully Convolutional Network) or DeepLab.

You may refer to the following MATLAB Documentation for the implementation of the architectures.

U-Net

https://www.mathworks.com/help/vision/ref/unetlayers.html

FCN

https://www.mathworks.com/help/vision/ref/fcnlayers.html

DeepLab

https://www.mathworks.com/help/vision/ref/deeplabv3pluslayers.html

 

For determining the orientation of the fibre you may refer to the Region and Image Properties from the MATLAB Documentation specifically the Orientation section:

https://www.mathworks.com/help/images/pixel-values-and-image-statistics.html?s_tid=CRUX_lftnav

 

I hope this helps!

Thank You

Maneet Bagga

[François]

Hi Maneet,

Thank you very much for your answer. I am indeed considering a neural network for the fibre segmentation, but I was wondering if there was a more 'simpler' approach as I don't have a lot of data (2 images in 30720x16016 .ndpi format).

I will anyway try what you suggested and come back to you !

Thank you again,

François Fournier

[Maneet Kaur Bagga]

Hi Francois,

For a smaller dataset, you can try these two alternate approaches:

1. Converting the image to a gray scale image, then applying a thresholding technique(global or adaptive) followed by applying morphological operations like erosion("imerode"), dilation("imdilate"), opening("imopen") or closing("imclose") to enhance the fibre structure and separate them from the backgorund.
2. Construct a graph representation of the image using the "graydiffweight" function to compute the weights between neighboring pixels based on their intensity differences. Apply graph-based segmentation algorithms like normalized cuts (nCut) or minimum spanning trees (graywatershed) to segment the fibers.