Which segmentation methods are preferable?

Hi friends!I am writing a paper on segmentation methods.I need to apply the segmentation technique to a radiographic image. The size of the image is too large(5000*7000). Can you suggest me some methods so that I can get the result fastly?

Answer 1

[Image Analyst]

imresize() will make the image smaller, if that's what you want.

[Mounika M]

yeah.it works.But now the problem is that I need to calculate mean,variance,psnr,etc.. I resized the image and found out results for thresholding,region growing and k means. For all these I got same values(may be because of same size) and got different values for my proposed algorithm as I didn't resized the image in this case.Can I submit the results this way?

[Image Analyst]

I guess. But I don't know why you think you needed to resize the image in the first place. For many things, resizing will have no, or little, effect on the measurement.

[Mounika M]

TO run the algorithm with image size of around 8000*6000 would be a hectic job right.So i chose to resize the image.

[Image Analyst]

You have to access every quad of 4 pixels in that to do bicubic interpolation, so I'm not sure that resizing would speed it up that much over what you'd be doing, but I guess you've timed it so you'd know. I deal with some images that are 9000 by 7000 pixels and they only take a few seconds to process.

[Mounika M]

I have applied image that has 8000 by 6000 pixels for region growing algorithm.I am attaching the code here.Its taking more than 2 days to run that.

http://www.mathworks.in/matlabcentral/fileexchange/19084-region-growing

[Image Analyst]

Where did you post your image? How do I know that you even need region growing instead of more traditional and common methods like thresholding and connected components labeling? Maybe you just think you need region growing because you see a mass of similarly colored pixels and aren't familiar with image processing and the methods and options that are available.

[Mounika M]

It is not that I require region growing in my work. I completed segmentation by my own algorithm and in my paper I just wanna show the difference between the outputs obtained through the region growing and my method. Now I need to show that my method is preferable than region growing, etc.. My prof suggested me to add some quality parameters to show that my method is efficient. What parameters can I include?

[Image Analyst]

Well I guess the time to complete it would be the main thing you'd want to show. You'd also want to show how accurate the two algorithms are compared to a physician's outline of the region.

[Mounika M]

I am sorry to say that I could not understand your second suggestion. Can you please elaborate. Will psnr, snr, etc..be useful as segmentation quality parameters ?

[Image Analyst]

You need to have some "ground truth" - in other words, the known, true, accurate segmentation. What/who can you trust to give you the ground truth? Would you trust a physician if they outlined the region and said "I know for a fact that this is the most accurate boundary for this tissue"? If so, then that's your ground truth and you need to compare all your algorithms against that.

Also, please look at this: http://sve.loni.ucla.edu/instructions/metrics/?wscr=1920x1200 and study up on ROC curves ( http://en.wikipedia.org/wiki/Receiver_operating_characteristic

[Mounika M]

Thank you so much