This is one of the simplest voxel similarity measures involving subtracting the two images and computing the mean sum of squares of this difference (SSD) image in the region of overlap. For N voxels in the overlap domain Q.TA>B this is given by
It can be shown that this measure is optimal when two measures differ only by Gaussian noise . Although we are usually interested in finding differences between the images, these are often so small that this measure remains the most effective. Image noise may not have a Gaussian distribution but this is unlikely to have a significant effect on performance. The measure is frequently used although it is sensitive to a small number of voxels having very different intensities - as might occur, for example, in contrast-enhanced serial MR imaging or during a dynamic sequence of PET images. Using the sum of absolute differences (SAD) rather than the sum of squared differences can reduce the effect of outliers.
Was this article helpful?