Specificity:
1.0000 -0.9802
-0.9802 1.0000
Generalisation ability:
1.0000 -0.9950
-0.9950 1.0000
Some comments:
1. Should \beta_l also appear in equation (1)?
It does.
2. Is the shuffle distance the best way to measure difference between two images? Why not use the overlap again?
We can't. There is a misconception here.
3. It's quite often that there are illumination changes among images to be aligned. How these metrics cope with in this situation?
The model is normalising intensities; this works fine in practice; the image set may have been pre-processed to balance intensities.
Comments to Authors:
A mostly clear paper.
However, the description in the text of the shuffle distance seems to contradict the caption of Figure 2 where the mean is not mentioned. The visual quality of Figures 4 and 7 is rather poor: I would suggest the use of a vector graphics format rather than bitmap. There is an extra comma after "that" in the last sentence of the conclusion.
Two more technical questions. Doesn't the shuffle distance (by the shuffling itself) mean that the non-rigid registration is not properly evaluated? In other words, local non-rigid deformations could be hidden by the shuffling, especially for large values of the radius. Is that a fair comment?
In Section 4, you have "the specificity obtained for the two groupwise methods is significantly better [...] implying better registration...". Was the registration indeed better? Can this be verified? After all, you are proposing a method to assess the registration so it would be nice to assess the performance of the assessment ;-)
TMI overlap experiments with MGH dataset