Segmenting bones using Neural Network

I have a database of hand X-ray images and I am trying to segment their bones. Some of the images have very poor contrast. I tried usual segmenting approaches, such as thresholding and using morphological filters, edge-based, or region growing methods; but none of these methods work for most of my images correctly and they cannot segment bones precisely. I was thinking of trying another method, such as Neural Network, but I don't know how to apply it to my problem exactly? I thought I can segment the bones of the some of the images (training data) manually; Then, I would have the set of training data and desired outputs (segmented image); After this step, I can train my Neural Network with the training data and their desired output and then examine the trained NN with the rest of the images (test data). Does it seem correct?

I would be thankful if you could help me with it or if you have any other idea for solving this problem?

Answer 1

[Ahmet Cecen]

If you do enough of these by hand first (different extreme cases hopefully around 25), you can create a library of bones, as I imagine there is only a finite amount of shapes and orientation each bone can take.

You can then convolve each bone in the library on the image and train a threshold on them to identify each bone.

This is sort of impersonating a manual convolutional neural network.