A straight-forward solution is to threshold the output of RRF, which represents the confidence of the detection/localization.

Note that it is hard to detect the existence before localization. People usually determine the locations of regions that have the highest possibilities (among others) to be the object of interest, and then look at how confident these localizations are. This is the way it goes especially when spatial constraints (e.g. asymmetry) are incorporated in the detection/localization, as seems to be the case of your task.

Dear Xiaomeng,

Thank you very much for your valuable insight. 

I am sorry for not being able to capture your solution on thresholding the output. I would be grateful to you if you could elaborate a little more on that if you have a few minutes to spare.

I was also thinking of the same line assuming that detection should come after localization. But it sounded strange to me when I first thought of it. 

Thank you once again.

I read that paper you mentioned, and now understood your question. In this method, given an unknown region, the decision on whether it is the organ of interest not only depends on its visual appearance but also depends on its spatial distance from all other regions. This idea makes sense because the structure of organs of most patients have an almost identical spatial layout. For example, if we know the positions of heart and kidney, we can easily estimate the position of stomach even if we don't see it.

If my understanding is correct, your question is like this. Consider a patient who has all organs except for his left kidney. The method will wrongly detect the position, where the left kidney should usually be but actually is not for this patient, as the left kidney.

One solution is to simply add a classification step after the method in that paper. In the training step, given ground truth bounding boxes, we can obtain multiple training regions for each organ. For each organ A, build a binary classifier in which one class is "is A" and the other is "is not A". Here, you should use visual features only, and should not use spatial features any more. You can use a SVM or a random (classification) forest as the classifier.

In the testing step, first use the method in that paper to obtain organ positions and corresponding organ names. From each position, we obtain a region image. Let B denote the organ name for this position. We can simply put this region into the classifier of B, and obtain a discrete decision, "is B" or "is not B".

Dear Xiaomeng,

I am very grateful to you for spending your valuable time on reading the paper that I have mentioned. 

You are absolutely correct about my decision depending on the spatial distance from all other organs. And you are again spot on about my question. 

And thank you very much for the provided solution. I think it will provide me very good results. I will definitely be trying that out. 

Thank you once again.