Speed and Accuracy Tradeoffs

The single most effective strategy for improving the speed of intensity-based registration algorithms is to only compute the cost function for some subset of the total voxels in the images. Tenfold improvements in speed can be achieved with only modest decreases in registration accuracy by sparsely sampling the images [21]. Avoidance of overly stringent convergence criteria can also be quite helpful.

The accuracy of intensity-based registration is generally believed to be in the subvoxel range. It is difficult to define gold standards that allow true accuracy to be verified with this degree of certainty (see the chapter "Validation of Registration Accuracy"). Using internal consistency measures, we have shown that registration results obtained with high resolution MRI images are consistent with errors in the 75-150 ^m range in images with 1-millimeter voxels [21]. Errors in low-resolution PET images with 6.75-millimeter slice thickness are on the order of2 millimeters, which can also be considered subvoxel accuracy [21]. Smoothing the data [21] and editing the data to remove extraneous structures that might move slightly relative to the object of interest [10] appear to improve accuracy in at least some contexts.

Was this article helpful?

0 0

Post a comment