Confusion matrix to validate image segmentation

I have a segmented image of an xray image. I need to make a confusion matrix to validate that segmentation. Can anyone tell me please how can I apply a confusion matrix to validate image segmentation?

The origial image and te segmented image are shown below.

The first one is the original image and the second one is the segmented image.

[KALYAN ACHARJYA]

Do you have gold data (Ground Truth Results)?

[Image Analyst]

Plus (you probably already know this) you can't make a confusion matrix on just one image. You'd need at least two and preferably hundreds or thousands.

[Ishara Nipuni]

I already have a large set of ground truth images and their segmented images( actually two hundred of images). Can you help me on making a confusion matrix with those images?

And also I want to know how to make classes for the confusion matrix with pixel values.

Answer 1

[Image Analyst]

If you have the Statistics and Machine Learning Toolbox, you can use confusionmat() and confusionchart(). Of course you still have to do a bunch of experiments to determine true positive, true negatives, false positives, and false negatives. I assume you know how to do that. It's basically a list of whether you got the diagnosis right or wrong. You have to decide how to determine if the segmentation is right or wrong, like use the Sorensen-Dice metric to see how much of your segmentation overlaps a hand drawn segmentation by the radiologist.

[KALYAN ACHARJYA]

After calculates the all parameters read here

[Ishara Nipuni]

Thank you so much...but I can't understand that what kind of parameters that I can use to determine if the segmentation is correct or wrong. Can you please help me on it with the Matlab code?

[Ishara Nipuni]

And also I have a problem with what are the true positive and true negative values with the above equations. Please help me to solve this problem.

[Image Analyst]

What parameters do you have available for you to vary? The threshold and minimum blob size for example? So you pick one set of parameters and determine if the image is a true positive, true negative, false positive, or false negative. Do it on a bunch of images and determine the counts for those, so the sum of those 4 results will sum to 200 if you had 200 images. Then you vary the parameters and try again. Each time you vary the parameters, you will get a point on the ROC cuve and two values you can add to your confusion matrix.

ROC in Wikipedia

[Ishara Nipuni]

Can I use pixel counts of some specific plot points of the above image as segmentation parameters to build a confusion matrix ?

And I also like to know whether there are any other methods rather than confusion matrices to validate the segmentation. If there are any other method, please tell me.