Hello, I am trying to find the centroid in b/w joints for a hand x-ray, i have been trying different approaches but couldn't find it for all the joints in one attempt

Any help would be greatly appreciated, I am tying to find the centroid in between all the joints of the fingers in the human hand x-ray image. I need a centroid in between every joint for fingers. Trying different approaches but I am still a newbie to this image proccesing tools. if not the code, I could take steps to find the centroid of every joints as an answer as well. Thank you in advance. example link to the image :- 2945667385191d66440188d5436a8a_gallery.jpeg (387×442) (radiopaedia.org)

[Image Analyst]

It's hard to tell where one bone ends and the other starts since they overlap. For what reason do you need to know the centroids?

See papers here:

20.7.4.5 Medical Applications, Wrist Joints

[mahesh]

Hello Thank you for the resposne, i dont need an accurate centroid for all the joints but i needed to be close at least, we have been assigned a group project, in which we have been given all the coordinates for the centroids and images to train a deeplearning model but the last objective of the project is that we need to find the centroids by our own and compare it to the given centroid(with the coordinates)of the image and show how much accuratly we were able to find the centroids by using edge detection and morphological operations and.... !

Any help would be greatly appreciated as we all are stuck on finding it.

Thank you!

Answer 1

[Image Analyst]

See attached program where you can manually place a point at each centroid. This will establish "ground truth" for the points, if they can be considered so, like if a radiologist marked them.

A traditional edge detection and morphological algorithm could range from excceedingly bad to fairly good. Some of the papers presumably had good algorithms - that's why they published them - and they came from dates before Deep Learning was invented. So I'm not sure how good of a traditional algorithm you want to compare your deep learning method to. Of course the published algorithms will be way better than anything you or I could come up with because they worked months or years on their algorithms.

[mahesh]

Hello, Thank you again for taking time. we need to show that we could succesfully implement all the concepts that have been thought in our project but locating manually isn't an option that would help us here, i need to find tha centroid or at least try to find it as close as possible but not all the images have overlaped joints because that dataset given to us is of arthritis patients. and we needed to find something like in the image attaced.

[Image Analyst]

But then you have no ground truth whatsoever and you cannot train a deep learning model. If you use a traditional image processing method (edges, morphology, etc.) then it gives some estimation of where the ground truth for the joint separations or bone centroids is, but your assumption is that those are no good and your deep learning model is better. Thus you cannot use the traditional method as ground truth, because your assumption is that it is not ground truth.

So, since traditional methods cannot be used as ground truth (because then by definition of being ground truth they will ALWAYS outperform your DL model), and you don't want to use manual location, then how exactly do you plan on getting ground truth? I know of no other way, unless there is some other trusted DL model that you want to use to define ground truth.

[mahesh]

Hello thank you for responding again, the project is a two parter. One is to train a DL model. For that we have got the corresponding coordinates for every centroid in a text file. So before feeding it to our model we need to write a function that can correctly locate those coordinates in an image by reading it from the text file, this part is done. The second part is, we need to write an algorithm using morphological, binarizaton and edge detection methods to find those centroids as closely accurate as we could but not necessarily very precise and compare it with the original image that has coordinates.

Thank you again for taking your time

[Image Analyst]

For the first part, how did you get those ground truth centroids in that text file? I'd guess it was by someone manually locating them, not by some automatic algorithm, because that is the second part of your task which you said you still have yet to do. Or someone before you may have trained a deep learning model and used that (though they'd still need manually located ground truth coordinates to train that model).

For that second part, I recommended some papers to review that should do the job pretty well. Did you look at any of those papers? If not, why not? Or you can just take a stab at it yourself for less reliable results.