How to use neural network to perform image segmentation?

I have a clinical need to perform lung segmentation on AP and PA chest radiographs. Using conventional segmentation techniques (thresholding, etc.) the adult lung is fairly easy to segment using automated methods. However, pediatric lungs are exceedingly difficult for a variety of reasons. My thought is to train a neural network to perform the task of lung identification using a set of manually segmented masks for training. Given that I know very little about neural networks, is there anyone who can help get me started by pointing me to some existing code, text, or other resource? Thanks in advance.

Example: Original Image

Example: Segmented Lung Mask

Answer 1

[Gautham Sholingar]

I understand that neural networks might seem like a useful avenue to try. Neural networks are useful for object detection and classification but segmentation might prove difficult especially since the regions change for every image. You will need a very large dataset and a custom neural network to make some progress in this area.

As of this stage it will be useful to understand the differences between adult and pediatric X-rays and how they affect the segmentation.

Could you explain why simple morphological segmentation is not yielding meaningful results for pediatric X-rays when it can yield good results for adult lungs?

It would also be useful to try the Image Segmenter App which is useful in such applications: http://www.mathworks.com/help/images/ref/imagesegmenter-app.html

[Jered Wells]

Thank you for your response. Pediatric lungs tend to be lower contrast and the images are subject to worse geometric (i.e. positioning) problems compared to the typical adult image. The small volume of the pediatric lung is also problematic, especially if disease is present. My thought is that a neural network might be able to take the position, size, shape, intensity, etc. and do a better job of figuring out where the lung is in an image compared to simply applying morphological operations (which I've done to an exhaustive extent already). I've even used a fuzzy c-means methodology that has yielded generally poor results thus far. Again, any help in the direction of setting up a neural network-type segmentation is much appreciated.

[Doga Tekin]

Hello Jered, I am kind of having the same struggle. I was wondering if you found out anything useful about setting up a neural network-type segmentation application in MATLAB since the time you asked this question. It would both greatly help me and any others trying to achieve the same thing.