Hello everyone, I have one question regarding image registration.

Can anyone help me how to calculate recall and precision value for registered image? I have two images that have been registered and currently looking for quantitative measurement method which is recall and precision. I also have the ground truth for those images. Please help, thank you.

Answer 1

[Image Analyst]

How about the sum of the absolute differences using imabsdiff()?

[nurfarisya zulkafli]

I am sorry sir, I dont get your point. Let say I have registered image and ground truth image, how can I calculate recall value (recall=tp/(tp+fn)) ? How to compute true positive(tn) and false negative(fn)? The images is in uint8 format.

please help, thankyou. I already provide the images for your reference.

[Image Analyst]

Did you look up the help for imabsdiff()? OK, let me talk you through it. Let's say you have two images that were perfectly registered. If you subtracted them, what do you think you'd get? You'd think you should get zero for every pixel, right? Because there is no difference in pixel values so they subtract away to zero. Now let's say there was some misalignment in the images. If you subtract those, what do you think you'd get? You'd get some non-zero values, right? And the more pixels that are misligned, or the greater the intensity or color difference the more the subtracted values would be. So that's the essence of imabsdiff() - it gives you an image where the value of the image is the delta intensity between the two images, on a pixel-by-pixel basis. So if you summed that difference image, or took it's mean, that would be a metric that characterizes how much misalignment between the images there is. Does that make sense?

What does not make sense (I think) is asking for a true positive or true negative when that has something to do with alignment. A single image does not have a TP or FP rate. How are you defining it? Is it if the mean absolute difference is less than something, meaning the alignment was fairly successful? Is it whether your algorithm identified disease or no disease in the image? Some other definition??? To build up an ROC curve, you have to change something in your algorithm and then determine whether it was correct or incorrect according to some criteria. So then you test a bunch of images and determine the TP and FP rate, and then that gets plotted as ONE point on the ROC curve. Now you change your parameters in your algorithm and do the test again. Now you get another point. Repeat that process and get more and more points along the ROC curve until you have as many as you think you need.