Hi Faiza,

If you find it, please do let us all know ;-)

Ground truth, in segmentation, refers to the "right answer", as you likely know. If you have a software which tells you the "right answer, how would you know? Perhaps because you segmented it manually yourself and are 110% confident in your interpretation? This is what most folks do, and thus ground truths are often estimated by a human being, such that the output of automated algorithms may be compared to this "right answer" for validation. However, I tend to have issue with this approach since typically, the very point of automated algorithms, beyond their speed, is to avoid subjectivity, which humans introduce, so how then can you validate an algorithm against a subjective "answer"?

In reality, the important question to ask of your data is not "what is the correct segmentation?", but rather "what segmentation accurately answers the question which prompted me to collect the data in the first place?" Since questions are typically answered by evaluating trends in data, it is often more important that you exercise rules-based , rather than subjective segmentation techniques across multiple datasets. This is what automated techniques allow. They may not produce the "right answer" everytime, but their error (however you define this) is at least systematic, and based on objective metrics, such that the trends you observe are accurate and un-biased.

With all this said, I'd like to mention a new software package we have just released, that we as scientists and end-users, not sales folks, are very excited about. It is available at http://MIPAR.us and offers versatile tools for 2D and 3D segmentation which incorporate and permit these rules-based approaches. I am always here to make you a segmentation recipe file for your data and am happy to help in any way I can. Best of luck in your research.

Cheers,

John

Ground truths are usually created by experts, practitioners, and/or radiologists for the image/volume under study not by software. Ground truth are considered as the perfect or perfect most accurate segmentation. Then you use the ground truth to compare the accuracy of  your segmentation with it, mostly using Dice coefficient, Jaccard similarity or Hausdorff distance.

You better use standard dataset or publicly available database to avoid this obstacle as if you tried to make the ground truth is very time consuming and definitely requires a lot of experience which mostly not available for junior researchers.

Good Luck 

Thank you Ahmed.